Apparatus and method for scheduling internet of things devices

Abstract

A scheduler based on proportional fairness is not suitable for high-entropy multi-request distribution, and the high-entropy multi-request distribution is expected by high-density deployment of Internet of Things equipment expected by a 6G network. To solve this problem, a method and apparatus for scheduling uplink transmissions based on contribution to downstream tasks is provided. And based on the contribution index of each device, selecting from the candidate device set to schedule a plurality of devices for uplink transmission. The contribution indicator for each device is related to a downstream task in a wireless communication network and indicates how much the device can successfully send information for the downstream task to the network and how much the information provided by the device is for the amount of information of the downstream task. The contribution index of the candidate device may be learned by machine learning using a deep neural network.

Description

Technical Field

[0001] The present disclosure relates generally to wireless communications and, in particular embodiments, to methods and apparatus for scheduling multiple devices in a wireless communication network. Background Art

[0002] In typical modern wireless communication systems such as wide band code division multiple access (WCDMA), long-term evolution (LTE), 5th Generation (5G), Wi-Fi, etc., many electronic devices (EDs) (also referred to as clients, terminals, user equipment (UE), mobile stations, etc.) can be connected to base stations (BSs) (also referred to as base transceiver stations (BTSs), Node-Bs, eNodeBs, gNBs, access points (APs), transmission points (TPs), etc.) over the air or are associated with the base stations. As the number and density of EDs increase, it becomes challenging to support high-quality wireless communications using traditional wireless systems.

[0003] Machine-to-machine (M2M) communication can be a type of high-density wireless communication. M2M communication is a technology that implements a network for collecting information from devices (e.g., sensors, smart meters, Internet of Things (IoT) devices, and / or other low-end devices) that are typically deployed on a large scale and densely, and for sending the information captured by these devices to other applications in the network. M2M networks can be wired or wireless and may have a relatively large geographic distribution (e.g., across a country or across the world). M2M communication typically does not involve direct human intervention to collect information.

[0004] 5G New Radio (NR) systems include capabilities to support massive machine type communications (mMTC), which connect large numbers (e.g., millions or billions) of IoT devices through wireless systems. It is expected that in the near future, the amount of M2M communications conducted over the air will exceed the amount of communications associated with humans. For example, the 6th Generation (6G) system is expected to connect more IoT devices than mobile phones. In 6G, high-density IoT deployments are expected to spawn many innovative applications, profoundly reshaping many industries and society. Some forecasts predict that deployment densities in 6G systems could reach 1km per 1km 2 10 9 It will be a challenge for 6G systems to support such high-density IoT deployments, where thousands or tens of thousands of IoT devices may simultaneously send their data back to the network over a shared wireless channel.

[0005] It is therefore desirable to provide a method for improving wireless communications, including improvements to accommodate and optimize the scheduling of large numbers of densely deployed IoT devices. Summary of the invention

[0006] According to the first broad aspect of the present disclosure, a method for scheduling uplink transmission in a wireless communication network is provided herein. The method according to the first broad aspect of the present disclosure may include selecting a first plurality of devices from a set of candidate devices to schedule for uplink transmission, wherein the selection of the first plurality of devices is based on a first contributingness metric of each device. The first contributingness metric of each device may be related to a first downstream task in the wireless communication network. For example, the first contributingness metric of a given device may indicate: i) to what extent the device can successfully send information for the first downstream task to the network; and ii) how much information the information provided by the device is for the first downstream task. The method according to the first broad aspect of the present disclosure may also include sending scheduling information indicating uplink wireless resources allocated to the first plurality of devices.

[0007] Providing contribution-based scheduling according to the first broad aspect of the present disclosure may have several advantages. For example, contribution-based scheduling may avoid scheduling the device with the smallest contribution, for example, due to its unfavorable observation location or due to its severe path loss or both, which avoids wasting of radio resources that may be allocated to the device with the smallest contribution when using traditional request-based proportional fairness, as discussed in further detail herein. In addition, since the contribution metric may be specific to a given downstream task, contribution-based scheduling may take into account the different contributions of devices between different tasks, for example, a device that contributes a lot to one task may be irrelevant to another task.

[0008] In some embodiments, for each device in the first plurality of devices, uplink wireless resources are allocated to the device based on the first contribution indicator of the device, thereby allocating the uplink wireless resources to the first plurality of devices. For example, a device having a first contribution indicator indicating a higher contribution to the first task can be allocated more uplink wireless resources than a device having a first contribution indicator indicating a lower contribution to the first task.

[0009] In some embodiments, the method according to the first broad aspect of the present disclosure also includes selecting a second plurality of devices from the set of candidate devices to be scheduled for uplink transmission. The selection of the second plurality of devices may be based on a second contribution indicator of each device. For example, the second contribution indicator of each device may be associated with a second downstream task in the wireless communication network that is different from the first downstream task. For example, the second contribution indicator may indicate: i) to what extent the device can successfully send information for the second downstream task to the network; and ii) how much information the information provided by the device is for the second downstream task. Then, scheduling information indicating the uplink wireless resources allocated to the second plurality of devices may be sent to the second plurality of devices.

[0010] In some embodiments, the first contribution metric of the candidate devices for the first downstream task is learned by using machine learning of a machine learning module. For example, the machine learning module may include a deep neural network (DNN), and the DNN is trained using raw test data received from at least one subset of the candidate devices as the ML module input and one or more parameters for the first downstream task as the ML module output to meet a training objective related to the first downstream task. For example, in some embodiments, the DNN may be configured as an autoencoder including at least two layers of neurons, wherein the first layer of the autoencoder is a linear fully-connected layer, the linear fully-connected layer includes K neurons, the K neurons have N inputs corresponding to a set of N candidate devices and K outputs, and each output of the K outputs of the first layer is a weighted linear combination of the N inputs, wherein once trained, the first layer of the autoencoder is configured to select an N-to-K selector of K inputs from the set of N inputs, where K < N. In such embodiments, one or more layers after the first layer of the autoencoder may be configured as a decoder to perform decoding of the first downstream task by using the K outputs from the first layer as the input to the decoder.

[0011] In some embodiments, the training of the autoencoder is based on stochastic gradient descent (SGD) backpropagation from the last layer of the autoencoder to the first layer to meet a training objective related to the first downstream task.

[0012] In some embodiments, the first plurality of devices are selected based on the weights of the first layer of the autoencoder after training for uplink transmission scheduling.

[0013] This is a nondeterministic polynomial (NP) problem for jointly optimizing the information quantity metric of the devices for a task and the conditions of their channel connections to the network. The advantage of using the DNN as described above is that, for example, the backpropagation training algorithm of SGD can propagate the task (i.e., the training objective) from the last layer to the first layer, which means that all neurons of the DNN work together to complete the task. The first layer will regulate the scheduler, and the remaining layers will fuse and process the incoming information from multiple devices. In addition, information about the path loss rate among the devices is embedded in the training and test data sets. Since it is data-driven, this channel path loss factor will be implicitly taken into account in the optimization of the DNN. The autoencoder (DNN)-based scheduling method disclosed herein provides a global optimization platform for joint optimization.

[0014] In some embodiments, the training of the autoencoder polarizes the weights in the first layer of the autoencoder such that for each of the K neurons in the first layer of the autoencoder, the output of the neuron is a weighted combination of the N inputs to the neuron, but only one of the N weights is close to a value of 1 and the remaining N-1 weights are close to a value of 0. For example, the training of the autoencoder can utilize a continuous relaxation of a discrete distribution (specific distribution) parameterized by a temperature parameter, wherein the temperature parameter decreases over the course of multiple training epochs such that the weights of the first layer of the autoencoder become increasingly polarized. In some such embodiments, for each of the K neurons in the first layer of the autoencoder, a candidate device corresponding to an input whose trained weight of the neuron is close to a value of 1 is considered to have been selected by the neuron. For example, in some embodiments, the number of neurons K in the first layer of the autoencoder can be equal to the K number of neurons of the first downstream task. min , where K min is a downstream task-specific value, and K of the first downstream task min Identifying, during training of the autoencoder for the first downstream task, and indicating a minimum number of neurons in the first layer that enables a training objective associated with the first downstream task to be met without requiring more than one of the neurons in the first layer to select any candidate device. For example, in some embodiments, K for the first downstream task may be determined during training of the autoencoder for the first downstream task by training multiple versions of the autoencoder using the same original test data as an input to the autoencoder and the same training objective associated with the first downstream task, but with different numbers of neurons in the first layer of the autoencoder. min .

[0015] In some embodiments, at least one of the candidate devices may be selected by more than one of the K neurons in the first layer. In some such embodiments, the first contribution index of each candidate device may be based on the number of times the candidate device was selected in the first layer of the autoencoder.

[0016] In some embodiments, the method according to the first broad aspect of the present disclosure also includes: dividing the candidate devices into multiple groups based on the number of times the candidate devices are selected in the first layer of the autoencoder, and the multiple groups include at least a primary group and a secondary group. For example, the candidate devices grouped into the primary group can be selected more times in the first layer of the autoencoder than the candidate devices grouped into the secondary group. In such embodiments, selecting the first plurality of devices to schedule for uplink transmission may include selecting devices of the primary group, and sending uplink scheduling information for the first plurality of devices may include sending primary uplink scheduling information for devices of the primary group. In some such embodiments, each of the candidate devices grouped into the secondary group can be selected at least once in the first layer of the autoencoder.

[0017] In some embodiments, the method according to the first broad aspect of the present disclosure further comprises receiving an uplink transmission from a device in the primary group of devices according to the primary uplink scheduling information. In some such embodiments, the uplink transmission received from the device in the primary group can be used as an input to a trained decoder to perform decoding of the first downstream task.

[0018] In some embodiments, the method according to the first broad aspect of the present disclosure further includes: determining one or more confidence indicators based on decoding of the first downstream task that uses uplink transmissions received from devices of the primary group as input to the trained decoder. In some such embodiments, the method may also include determining whether to send secondary uplink scheduling information for devices of the secondary group based on the one or more confidence indicators. For example, determining whether to send secondary uplink scheduling information for devices of the secondary group may include determining not to send secondary uplink scheduling information for devices of the secondary group after determining that the one or more confidence indicators indicate that the result of decoding of the first downstream task that uses uplink transmissions received from devices of the primary group as input to the trained decoder has sufficient confidence.

[0019] In some embodiments, the method according to the first broad aspect of the present disclosure further includes: after determining that the one or more confidence indicators indicate that the result of decoding of the first downstream task using the uplink transmission received from the device of the primary group as the input of the trained decoder has insufficient confidence, sending secondary uplink scheduling information for the devices of the secondary group, the secondary uplink scheduling information indicating the uplink wireless resources allocated to the device for each device in the secondary group. In some such embodiments, the method may also include: receiving an uplink transmission from a device in the secondary group according to the secondary uplink scheduling information; using the uplink transmission received from the device of the primary group and the uplink transmission received from the device of the secondary group as the input of the trained decoder to perform decoding of the first downstream task. In some embodiments, determining one or more confidence indicators may include determining one or more softmax values.

[0020] Corresponding apparatus and devices for performing the method are disclosed.

[0021] For example, according to another aspect of the present disclosure, a network device is provided, which includes a processor and a memory storing processor-executable instructions, and the processor-executable instructions, when executed, cause the processor to perform the method according to the first broad aspect of the present disclosure mentioned above.

[0022] According to other aspects of the present disclosure, a device is provided, including one or more units for implementing any method aspects disclosed in the present disclosure. The term "unit" is used in a broad sense and can be referred to by any of a variety of names, including, for example, a module, a component, an element, a member, etc. These units can be implemented using hardware, software, firmware, or any combination thereof. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Reference will now be made, by way of example only, to the accompanying drawings which show exemplary embodiments of the present application, in which:

[0024] Figure 1 is a simplified schematic diagram of a communication system according to an example;

[0025] Figure 2 Another example of a communication system is shown;

[0026] Figure 3 Examples of electronic devices (EDs), terrestrial transmit and receive points (T-TRPs), and non-terrestrial transmit and receive points (NT-TRPs) are shown;

[0027] Figure 4 Exemplary units or modules in the device are shown;

[0028] Figure 5 illustrates four EDs communicating with a network device in a communication system according to one embodiment;

[0029] Fig. 6A An example of a neural network having multiple neuron layers according to one embodiment is shown;

[0030] Figure 6B shows an example of a neuron that may be used as a building block of a neural network according to one embodiment;

[0031] Figure 7 An example of a neural network configured as an autoencoder according to one embodiment is shown;

[0032] Figure 8 shows the entropy of a two-dimensional (2-device) request distribution;

[0033] Fig. 9 An example of an Internet of Things (IoT) deployment scenario is shown, where multiple IoT devices are measuring or observing a common object;

[0034] Fig.10 Shown in Fig. 9 The potential impact of different channel conditions in IoT deployment scenarios;

[0035] Fig.11A and Fig. 11B Shown in Fig. 9 Examples of different subsets of devices that can be scheduled for classification and detection tasks in IoT deployment scenarios;

[0036] Fig.12 Shows Fig.10 An example of an IoT deployment where several devices are experiencing high erasure channel conditions and therefore may not be able to successfully send data to the network;

[0037] Fig.13 An example of a DNN configured as an autoencoder according to one embodiment is shown, the DNN including a first scheduling layer acting as a scheduler and a plurality of decoding layers acting as fusers and processors for information received from devices selected by the scheduling layer;

[0038] Fig.14 shows an example of epoch-by-epoch cool-down during training according to one embodiment;

[0039] Fig.15shows that according to one embodiment, when the output dimension K of the scheduling layer is greater than K of a given task min Examples of duplicate selection of contributing devices that may occur when

[0040] Fig.16 It shows that the output dimension K of the scheduling layer according to one embodiment is much larger than K min Scenarios and examples of ranking of contributing devices that may be accomplished based on the number of times each contributing device is selected by the scheduling layer;

[0041] Fig.17 An example is shown according to one embodiment, where contributing devices have been grouped based on their contribution ranking, and the groups are then incrementally scheduled as needed to confidently complete downstream tasks;

[0042] Fig.18 An example of a less contributing device acting as a relay for a more contributing device experiencing poor channel conditions is shown according to one embodiment;

[0043] Fig.19 An example is shown of 784 sensors or devices uniformly deployed in a 28x28 array to observe an area according to one embodiment.

[0044] Fig. 20 shows the configuration of a specific encoder and decoder of size K according to one embodiment, the decoder being used to simulate the MNIST image dataset as input data. Fig.19 deployment scenarios.

[0045] Fig.21 According to an embodiment Fig. 20 The implementation of the decoder.

[0046] Fig. 22 Show MNIST handwritten digit input image;

[0047] Fig.23 A graph showing simulated accuracy percentage versus erasure probability percentage for a 10-class classification task for different values ​​of encoder size output K according to one embodiment;

[0048] Fig.24 shows the reconstructed MNIST images for different erasure probability percentages, K=128 / 256, according to one embodiment;

[0049] Fig.25 A graph showing simulated accuracy percentage versus erasure probability percentage for a reconstruction task for different values ​​of encoder size output K according to one embodiment;

[0050] Fig.26A graph showing simulated accuracy percentage versus erasure probability percentage for a detection task for different values ​​of encoder size output K according to one embodiment;

[0051] Fig. 27 shows different values ​​of encoder size output K and 0% erasure probability percentage for a classification task, for a selected set of devices in a 28x28 array, according to one embodiment;

[0052] Fig.28 shows different values ​​of encoder size output K and 50% erasure probability percentage for a classification task, for a selected set of devices in a 28x28 array, according to one embodiment;

[0053] Fig.29 shows different values ​​of encoder size output K and 0% erasure probability percentage for a reconstruction task, for a selected set of devices in a 28x28 array, according to one embodiment;

[0054] Fig.30 shows different values ​​of encoder size output K and 0% erasure probability percentage for a detection task, for a selected set of devices in a 28x28 array, according to one embodiment;

[0055] Fig.31 shows an example of a non-uniform erasure channel used in a second set of simulations according to one embodiment;

[0056] Fig.32 Shown in Fig.31 An example of transmission results over a non-uniform erasure channel is shown;

[0057] Fig.33 10-class classification task for different values ​​of erasure percentage probability according to one embodiment is shown in FIG. Fig.31 Plot of simulated accuracy percentage vs. encoder size output K on a non-uniformly erased channel shown;

[0058] Fig.34 According to one embodiment, the classification task Fig.31 The encoder size output K on the non-uniform erasure channel is shown for different values, for a selected set of devices in a 28x28 array;

[0059] Fig.35 The reconstruction tasks for different values ​​of the probability of erasure percentage according to one embodiment are shown in FIG. Fig.31 Plot of simulated accuracy percentage vs. encoder size output K on a non-uniformly erased channel shown;

[0060] Fig.36 According to one embodiment, the reconstruction task is shown in Fig.31The encoder size output K on the non-uniform erasure channel is shown for different values, for a selected set of devices in a 28x28 array;

[0061] Fig.37 According to one embodiment, Fig.31 Reconstructed MNIST images with K=32 / 128 / 256 / 512 on the non-uniformly erased channels shown;

[0062] Fig.38 FIG. 1 shows a detection task for different values ​​of the erase percentage probability according to one embodiment. Fig.31 Plot of simulated accuracy percentage vs. encoder size output K on a non-uniformly erased channel shown;

[0063] Fig.39 According to one embodiment, the detection task is Fig.31 The encoder size output K on the non-uniform erasure channel is shown for different values, for a selected set of devices in a 28x28 array;

[0064] Fig.40 A graph showing simulated accuracy percentage for a 10-class classification task under HARQ training versus encoder size output K and 0% erasure probability percentage according to one embodiment;

[0065] Fig.41 shows a selected set of devices in a 28x28 array for different values ​​of encoder size output K and 0% erasure probability percentage under HARQ training for a classification task according to one embodiment;

[0066] Fig.42 A graph showing simulated accuracy percentage of a reconstruction task under HARQ training versus encoder size output K and 0% erasure probability percentage according to one embodiment;

[0067] Fig.43 shows a selected set of devices in a 28x28 array for different values ​​of encoder size output K and 0% erasure probability percentage for a reconstruction task under HARQ training according to one embodiment;

[0068] Fig.44 A graph showing simulated accuracy percentage for a detection task under HARQ training versus encoder size output K and 0% erasure probability percentage according to one embodiment; and

[0069] Fig.45 A selected set of devices in a 28x28 array are shown for different values ​​of encoder size output K and 0% erasure probability percentage for a detection task under HARQ training according to one embodiment.

[0070] The same reference numerals may be used in different drawings to denote the same components. DETAILED DESCRIPTION

[0071] For illustrative purposes, specific exemplary embodiments are explained in detail below with reference to the accompanying drawings.

[0072] Exemplary Communication Systems and Devices

[0073] refer to Figure 1 , a simplified schematic diagram of a communication system is provided as an illustrative example but not limiting. The communication system 100 includes a radio access network 120. The radio access network 120 may be a next generation (e.g., sixth generation (6G) or higher) radio access network or a traditional (e.g., 5G, 4G, 3G or 2G) radio access network. One or more communication electronic devices (electric devices, ED) 110a to 120j (collectively referred to as 110) may be interconnected with each other or connected to one or more network nodes (170a, 170b, collectively referred to as 170) in the radio access network 120. The core network 130 may be part of the communication system and may be dependent on or independent of the radio access technology used in the communication system 100. In addition, the communication system 100 includes a public switched telephone network (PSTN) 140, the Internet 150, and other networks 160.

[0074] Figure 2 An exemplary communication system 100 is shown. Typically, the communication system 100 enables multiple wireless or wired elements to transmit data and other content. The purpose of the communication system 100 may be to provide content such as voice, data, video and / or text by broadcasting, multicasting and unicasting, etc. The communication system 100 may operate by sharing resources (e.g., carrier spectrum bandwidth) between its constituent elements. The communication system 100 may include a terrestrial communication system and / or a non-terrestrial communication system. The communication system 100 may provide a wide range of communication services and applications (e.g., earth monitoring, remote sensing, passive sensing and positioning, navigation and tracking, autonomous transmission and mobility, etc.). The communication system 100 may provide high availability and robustness through the joint operation of a terrestrial communication system and a non-terrestrial communication system. For example, integrating a non-terrestrial communication system (or its components) into a terrestrial communication system may produce a heterogeneous network that is considered to include multiple layers. Compared with traditional communication networks, heterogeneous networks may achieve better overall performance through efficient multi-link joint operation, more flexible function sharing, and faster physical layer link switching between terrestrial networks and non-terrestrial networks.

[0075] The terrestrial communication system and the non-terrestrial communication system can be considered as subsystems of the communication system. In the example shown, the communication system 100 includes electronic devices (ED) 110a to 110d (generally referred to as ED 110), radio access networks (RAN) 120a to 120b, non-terrestrial communication networks 120c, core networks 130, public switched telephone networks (PSTN) 140, Internet 150 and other networks 160. RAN 120a to 120b include corresponding base stations (BS) 170a to 170b, which can be generally referred to as terrestrial transmit and receive points (T-TRP) 170a to 170b. Non-terrestrial communication network 120c includes access node 120c, which can be generally referred to as non-terrestrial transmit and receive point (NT-TRP) 172.

[0076] Any ED 110 may alternatively or additionally be used to connect, access, or communicate with any other T-TRP 170a to 170b, NT-TRP 172, Internet 150, core network 130, PSTN 140, other network 160, or any combination thereof. In some examples, ED 110a may communicate uplink and / or downlink transmissions with T-TRP 170a via interface 190a. In some examples, ED 110a, 110b, and 110d may also communicate directly with each other via one or more sidelink air interfaces 190b. In some examples, ED 110d may communicate uplink and / or downlink transmissions with NT-TRP 172 via interface 190c.

[0077] The air interfaces 190a and 190b may use similar communication technologies, such as any applicable wireless access technology. For example, the communication system 100 may implement one or more channel access methods in the air interfaces 190a and 190b, such as code division multiple access (CDMA), time division multiple access (TDMA), frequency division multiple access (FDMA), orthogonal FDMA (OFDMA), or single-carrier FDMA (SC-FDMA). The air interfaces 190a and 190b may utilize other high-dimensional signal spaces, which may include a combination of orthogonal dimensions and / or non-orthogonal dimensions.

[0078] Air interface 190c may enable communication between ED 110d and one or more NT-TRPs 172 via a wireless link or a simple link. For some examples, the link is a dedicated connection for unicast transmission, a connection for broadcast transmission, or a connection between a group of EDs and one or more NT-TRPs for multicast transmission.

[0079] The RANs 120a and 120b communicate with the core network 130 to provide various services, such as voice, data, and other services, to the EDs 110a, 110b, and 110c. The RANs 120a and 120b and / or the core network 130 may communicate directly or indirectly with one or more other RANs (not shown), which may or may not be directly served by the core network 130 and may or may not employ the same radio access technology as the RANs 120a, 120b, or both. The core network 130 may also serve as a gateway access between (i) the RANs 120a and 120b, or the EDs 110a, 110b, and 110c, or both, and (ii) other networks, such as the PSTN 140, the Internet 150, and other networks 160. In addition, some or all of the EDs 110a, 110b, and 110c may include functionality for communicating with different wireless networks over different radio links using different radio technologies and / or protocols. Instead of (or in addition to) wireless communication, EDs 110a, 110b, and 110c may also communicate with a service provider or switch (not shown) and with the Internet 150 through a wired communication channel. PSTN 140 may include a circuit-switched telephone network for providing plain old telephone service (POTS). Internet 150 may include a computer network, a subnet (intranet), or both, and include protocols such as Internet Protocol (IP), Transmission Control Protocol (TCP), User Datagram Protocol (UDP), etc. EDs 110a, 110b, and 110c may be multi-mode devices capable of operating according to a variety of wireless access technologies and include a plurality of transceivers required to support these technologies.

[0080] Figure 3Another example of ED 110 and base stations 170a, 170b and / or 170c is shown. ED 110 is used to connect people, objects, machines, etc. ED 110 can be widely used in various scenarios, such as cellular communication, device-to-device (D2D), vehicle to everything (V2X), peer-to-peer (P2P), machine-to-machine (M2M), machine-type communication (MTC), Internet of Things (IOT), virtual reality (VR), augmented reality (AR), industrial control, autonomous driving, telemedicine, smart grid, smart furniture, smart office, smart wearable devices, smart transportation, smart city, drones, robots, remote sensing, passive sensing, positioning, navigation and tracking, autonomous delivery and movement, etc.

[0081] Each ED 110 represents any end-user device suitable for wireless operation, and may include (or may be referred to as): user equipment / device (UE), wireless transmit / receive unit (WTRU), mobile station, fixed or mobile user unit, cellular phone, station (STA), machine type communication (MTC) device, personal digital assistant (PDA), smartphone, laptop, computer, tablet, wireless sensor, consumer electronic device, smartbook, vehicle, car, truck, bus, train or IoT device, industrial equipment or devices in the aforementioned devices (e.g., communication module, modem or chip), etc. The next generation ED 110 may be referred to using other terms. Base stations 170a and 170b are T-TRPs and are referred to as T-TRP 170 hereinafter. Also in Figure 3 , the NT-TRP will be referred to hereinafter as NT-TRP 172. Each ED 110 connected to T-TRP 170 and / or NT-TRP 172 may be dynamically or semi-statically started (i.e., established, activated, or enabled), shut down (i.e., released, deactivated, or disabled), and / or configured in response to one or more of connection availability and connection necessity.

[0082] ED 110 includes a transmitter 201 and a receiver 203 coupled to one or more antennas 204. Only one antenna 204 is shown. Alternatively, one, some or all of the antennas may be panels. Transmitter 201 and receiver 203 may be integrated into, for example, a transceiver. The transceiver is used to modulate data or other content for transmission via at least one antenna 204 or a network interface controller (NIC). The transceiver is also used to demodulate data or other content received by at least one antenna 204. Each transceiver includes: any suitable structure for generating signals for wireless or wired transmission and / or for processing signals received by wireless or wired means. Each antenna 204 includes: any suitable structure for sending and / or receiving wireless signals or wired signals.

[0083] ED 110 includes at least one memory 208. The memory 208 stores instructions and data used, generated, or collected by ED 110. For example, the memory 208 may store software instructions or modules for implementing some or all of the functions and / or embodiments described herein and executed by the processing unit 210. Each memory 208 includes any suitable volatile and / or non-volatile storage and retrieval device. Any suitable type of memory may be used, such as random access memory (RAM), read only memory (ROM), hard disk, optical disk, subscriber identity module (SIM) card, memory stick, secure digital (SD) memory card, on-processor cache, etc.

[0084] ED 110 may also include one or more input / output devices (not shown) or interfaces (e.g., Figure 1 The input / output devices support interaction with users or other devices in the network. Each input / output device includes any suitable structure for providing information to or receiving information from a user, such as a speaker, microphone, keypad, keyboard, display or touch screen, including network interface communications.

[0085] ED 110 also includes a processor 210 for performing operations, including operations related to preparing uplink transmissions sent to NT-TRP 172 and / or T-TRP 170, processing downlink transmissions received from NT-TRP 172 and / or T-TRP 170, and processing sidelink transmissions with another ED 110. Processing operations related to preparing to send uplink transmissions may include operations such as encoding, modulation, transmit beamforming, and generating symbols for transmission. Processing operations related to processing downlink transmissions may include operations such as receive beamforming, demodulation, and decoding received symbols. According to an embodiment, the downlink transmission may be received by receiver 203 using receive beamforming, and processor 210 may extract signaling from the downlink transmission (e.g., by detecting and / or decoding the signaling). An example of signaling may be a reference signal sent by NT-TRP 172 and / or T-TRP 170. In some embodiments, the processor 276 implements transmit beamforming and / or receive beamforming based on an indication of a beam direction (e.g., beam angle information (BAI)) received from the T-TRP 170. In some embodiments, the processor 210 may perform operations related to network access (e.g., initial access) and / or downlink synchronization, such as operations related to detecting a synchronization sequence, decoding, and obtaining system information. In some embodiments, the processor 210 may perform channel estimation, for example, using a reference signal received from the NT-TRP 172 and / or the T-TRP 170.

[0086] Although not shown, the processor 210 may form a part of the transmitter 201 and / or the receiver 203. Although not shown, the memory 208 may form a part of the processor 210.

[0087] The processor 210 and the processing components of the transmitter 201 and the receiver 203 may each be implemented by the same or different one or more processors, which are used to execute instructions stored in a memory (e.g., the memory 208). Alternatively, the processor 210 and some or all of the processing components of the transmitter 201 and the receiver 203 may be implemented using a dedicated circuit such as a programmed field-programmable gate array (FPGA), a graphical processing unit (GPU), or an application-specific integrated circuit (ASIC).

[0088] T-TRP 170 may use other names in some implementations, such as base station, base transceiver station (BTS), wireless base station, network node, network device, network side device, sending / receiving node, Node B, evolved NodeB (eNodeB or eNB), home eNodeB, next generation NodeB (gNB), transmission point (TP), site controller, access point (AP), wireless router, relay station, remote radio head, ground node, ground network device, ground base station, base band unit (BBU), remote radio unit (RRU), active antenna unit (AAU), remote radio head (RRH), centralized unit (CU), distributed unit (DU), positioning node, etc. T-TRP 170 may be a macro BS, a micro BS, a relay node, a host node, etc. or a combination thereof. T-TRP 170 may refer to the above-mentioned device, or refer to a device in the above-mentioned device (eg, a communication module, a modem, or a chip).

[0089] In some embodiments, various parts of T-TRP 170 may be distributed. For example, some modules of T-TRP 170 may be located away from the device that houses the antenna of T-TRP 170, and may be coupled to the device that houses the antenna through a communication link (not shown) sometimes referred to as a fronthaul, such as a common public radio interface (CPRI). Therefore, in some embodiments, the term T-TRP 170 may also refer to modules on the network side that perform processing operations such as determining the location of ED 110, resource allocation (scheduling), message generation, and encoding / decoding, and are not necessarily part of the device that houses the antenna of T-TRP 170. Modules may also be coupled to other T-TRPs. In some embodiments, T-TRP 170 may actually be multiple T-TRPs that operate together to serve ED 110, for example, through coordinated multi-point transmission.

[0090] The T-TRP 170 includes at least one transmitter 252 and at least one receiver 254 coupled to one or more antennas 256. Only one antenna 256 is shown. Alternatively, one, some or all of the antennas may be panels. The transmitter 252 and the receiver 254 may be integrated into a transceiver. The T-TRP 170 also includes a processor 260 for performing operations related to preparing downlink transmissions sent to the ED 110, processing uplink transmissions received from the ED 110, preparing backhaul transmissions sent to the NT-TRP 172, and processing transmissions received from the NT-TRP 172 via the backhaul. Processing operations related to preparing to send downlink transmissions or backhaul transmissions may include operations such as encoding, modulation, precoding (e.g., MIMO precoding), transmit beamforming, and generating symbols for transmission. Processing operations related to processing received transmissions in the uplink or on the backhaul may include operations such as receive beamforming, demodulation, and decoding received symbols. The processor 260 may also perform operations related to network access (e.g., initial access) and / or downlink synchronization, such as generating the content of a synchronization signal block (SSB), generating system information, etc. In some embodiments, the processor 260 also generates a beam direction indication, such as a BAI, that the scheduler 253 may schedule for transmission. The processor 260 performs other network-side processing operations described herein, such as determining the location of the ED 110, determining the location of the deployed NT-TRP 172, etc. In some embodiments, the processor 260 may generate signaling, such as to configure one or more parameters of the ED 110 and / or one or more parameters of the NT-TRP 172, etc. Any signaling generated by the processor 260 is sent by the transmitter 252. Note that as used herein, "signaling" may also be referred to as control signaling. Dynamic signaling may be sent in a control channel such as a physical downlink control channel (PDCCH), while static or semi-static higher layer signaling may be included in a packet sent in a data channel such as a physical downlink shared channel (PDSCH).

[0091] The scheduler 253 may be coupled to the processor 260. The scheduler 253 may be included within or operate separately from the T-TRP 170, which may schedule uplink transmissions, downlink transmissions, and / or backhaul transmissions, including issuing scheduling grants and / or configuring scheduling-free (“configuration grants”) resources. The T-TRP 170 also includes a memory 258 for storing information and data. The memory 258 stores instructions and data used, generated, or collected by the T-TRP 170. For example, the memory 258 may store software instructions or modules for implementing some or all of the functionality and / or embodiments described herein and executed by the processor 260.

[0092] Although not shown, the processor 260 may form a part of the transmitter 252 and / or the receiver 254. In addition, although not shown, the processor 260 may implement the scheduler 253. Although not shown, the memory 258 may form a part of the processor 260.

[0093] The processor 260, the scheduler 253, and the processing components of the transmitter 252 and the receiver 254 may be implemented by the same or different one or more processors, respectively, which are used to execute instructions stored in a memory (e.g., the memory 258). Alternatively, some or all of the processing components of the processor 260, the scheduler 253, and the transmitter 252 and the receiver 254 may be implemented using a dedicated circuit such as an FPGA, a GPU, or an ASIC.

[0094] Although the NT-TRP 172 is shown as a drone by way of example only, the NT-TRP 172 may be implemented in any suitable non-ground form. In addition, in some implementations, the NT-TRP 172 may use other names, such as a non-ground node, a non-ground network device, or a non-ground base station. The NT-TRP 172 includes a transmitter 272 and a receiver 274 coupled to one or more antennas 280. Only one antenna 280 is shown. Alternatively, one, part, or all of the antennas may be panels. The transmitter 272 and the receiver 274 may be integrated into a transceiver. The NT-TRP 172 also includes a processor 276 for performing operations related to preparing downlink transmissions sent to the ED 110, processing uplink transmissions received from the ED 110, preparing backhaul transmissions sent to the T-TRP 170, and processing transmissions received from the T-TRP 170 via backhaul. Processing operations related to preparing to transmit downlink transmissions or backhaul transmissions may include operations such as encoding, modulation, precoding (e.g., MIMO precoding), transmit beamforming, and generating symbols for transmission. Processing operations related to processing received transmissions in the uplink or on the backhaul may include operations such as receive beamforming, demodulation, and decoding received symbols. In some embodiments, the processor 276 implements transmit beamforming and / or receive beamforming based on beam direction information (e.g., BAI) received from the T-TRP 170. In some embodiments, the processor 276 may generate signaling, for example, to configure one or more parameters of the ED 110, etc. In some embodiments, the NT-TRP 172 implements physical layer processing but does not implement higher-level functions, such as functions of the medium access control (MAC) layer or the radio link control (RLC) layer. Since this is just an example, more generally, the NT-TRP 172 may implement higher-level functions in addition to physical layer processing.

[0095] The NT-TRP 172 also includes a memory 278 for storing information and data. Although not shown, the processor 276 may form part of the transmitter 272 and / or the receiver 274. Although not shown, the memory 278 may form part of the processor 276.

[0096] Processor 276 and the processing components of transmitter 272 and receiver 274 may be implemented by the same or different one or more processors, respectively, that execute instructions stored in a memory (e.g., memory 278). Alternatively, processor 276 and some or all of the processing components of transmitter 272 and receiver 274 may be implemented using dedicated circuits such as a programmed FPGA, GPU, or ASIC. In some embodiments, NT-TRP 172 may actually be multiple NT-TRPs that operate together to serve ED 110, for example, via coordinated multi-point transmission.

[0097] Note that as used herein, "TRP" may refer to T-TRP or NT-TRP.

[0098] The T-TRP 170, NT-TRP 172, and / or ED 110 may include other components, but these components have been omitted for clarity.

[0099] according to Figure 4 , one or more steps in the embodiment method provided in this article can be performed by corresponding units or modules. Figure 4 Units or modules in a device, such as ED 110, T-TRP 170, or NT-TRP 172 are shown. For example, a signal can be sent by a sending unit or a sending module. For example, a signal can be sent by a sending unit or a sending module. A signal can be received by a receiving unit or a receiving module. The signal can be processed by a processing unit or a processing module. Other steps can be performed by an artificial intelligence (AI) module or a machine learning (ML) module. The corresponding units or modules can be implemented using hardware, one or more components or devices that execute software, or a combination thereof. For example, one or more units or modules can be integrated circuits such as programmed FPGAs, GPUs, or ASICs. It should be understood that if these modules are implemented, for example, using software executed by a processor, these modules can be retrieved by the processor in whole or in part as needed, retrieved individually or collectively for processing, retrieved in one or more instances, and these modules themselves may include instructions for further deployment and instantiation.

[0100] Additional details regarding ED 110, T-TRP 170, and NT-TRP 172 are known to those skilled in the art. Therefore, these details are omitted here.

[0101] Control signaling is discussed in some embodiments herein. Control signaling may sometimes be referred to alternatively as signaling, or control information, or configuration information, or configuration. In some cases, control signaling may be indicated dynamically, such as dynamically indicated in a physical layer in a control channel. An example of dynamically indicated control signaling is information sent in physical layer control signaling, such as downlink control information (DCI). Control signaling may sometimes be indicated semi-statically alternatively, such as semi-statically indicated in RRC signaling or in a MAC control element (CE). A dynamic indication may be an indication in a low layer, such as an indication in a physical layer / layer 1 signaling (e.g., in a DCI), rather than an indication in a high layer (e.g., rather than in an RRC signaling or MAC CE). A semi-static indication may be an indication in a semi-static signaling. Semi-static signaling used herein may refer to non-dynamic signaling, such as high layer signaling, RRC signaling, and / or MAC CE. Dynamic signaling used herein may refer to dynamic signaling, such as physical layer control signaling sent in a physical layer, such as a DCI.

[0102] An air interface typically includes multiple components and associated parameters that together specify how to send and / or receive transmissions over a wireless communication link between two or more communication devices. For example, an air interface may include one or more components that define a waveform, frame structure, multiple access scheme, protocol, coding scheme, and / or modulation scheme for transmitting information (e.g., data) over a wireless communication link. A wireless communication link may support a link between a radio access network and a user device (e.g., a "Uu" link), and / or a wireless communication link may support a link between devices, such as a link between two user devices (e.g., a "sidelink"), and / or a wireless communication link may support a link between a non-terrestrial (NT) communication network and a user equipment (UE). The following are some examples of the above components:

[0103] The waveform component can specify the shape and form of the signal being sent. Waveform options can include orthogonal multiple access waveforms and non-orthogonal multiple access waveforms. Non-limiting examples of such waveform options include orthogonal frequency division multiplexing (OFDM), filtered OFDM (f-OFDM), time-windowed OFDM, filter bank multicarrier (FBMC), universal filtered multicarrier (UFMC), generalized frequency division multiplexing (GFDM), wavelet packet modulation (WPM), faster than Nyquist (FTN) waveforms, and low peak to average power ratio waveforms (PAPR WF).

[0104] The frame structure component may specify the configuration of a frame or a group of frames. The frame structure component may indicate one or more of the time, frequency, pilot signature, code, or other parameters of a frame or a group of frames. More details of the frame structure will be discussed below.

[0105] The multiple access scheme component may specify multiple access technology options, including technologies that define how communicating devices share a common physical channel, such as time division multiple access (TDMA), frequency division multiple access (FDMA), code division multiple access (CDMA), single carrier frequency division multiple access (SC-FDMA), low density signature multicarrier code division multiple access (LDS-MC-CDMA), non-orthogonal multiple access (NOMA), pattern division multiple access (PDMA), lattice partition multiple access (LPMA), resource spread multiple access (RSMA), and sparse code multiple access (SCMA). In addition, multiple access technology options may include: scheduled access versus unscheduled access, also known as unlicensed access; non-orthogonal multiple access versus orthogonal multiple access, for example, through dedicated channel resources (e.g., not shared between multiple communication devices); contention-based shared channel resources versus non-contention-based shared channel resources; and cognitive radio-based access.

[0106] A hybrid automatic repeat request (HARQ) protocol component may specify how transmission and / or retransmission is performed. Non-limiting examples of transmission and / or retransmission mechanism options include those that specify a scheduled data pipe size, a signaling mechanism for transmission and / or retransmission, and a retransmission mechanism.

[0107] The coding and modulation components may specify how the information being sent may be encoded / decoded and modulated / demodulated for transmission / reception purposes. Coding may refer to methods of error detection and forward error correction. Non-limiting examples of coding options include turbo trellis codes, turbo product codes, fountain codes, low-density parity-check codes, and polar codes. Modulation may simply refer to constellations (including, for example, modulation techniques and orders), or more specifically to various types of advanced modulation methods, such as hierarchical modulation and low PAPR modulation.

[0108] In some embodiments, the air interface can be a "one-size-fits-all concept". For example, once the air interface is defined, the components within the air interface cannot be changed or adapted. In some implementations, only limited parameters or modes of the air interface can be configured, such as the cyclic prefix (CP) length or the multiple input multiple output (MIMO) mode. In some embodiments, the air interface design can provide a unified or flexible framework to support frequency bands below 6 GHz and above 6 GHz (e.g., millimeter wave) for authorized and unauthorized access. For example, the flexibility of the configurable air interface provided by the extensible parameter set and symbol duration can support transmission parameter optimization for different spectrum bands and different services / devices. For another example, a unified air interface can be self-contained in the frequency domain, and the self-contained design in the frequency domain can support more flexible radio access network (RAN) slicing by sharing channel resources between different services in frequency and time.

[0109] Frame structure

[0110] The frame structure is a feature of the wireless communication physical layer, which defines the time domain signal transmission structure, for example, to achieve the timing reference and timing alignment of the basic time domain transmission unit. Wireless communication between communication devices can be carried out on the time-frequency resources controlled by the frame structure. The frame structure can sometimes be referred to as the radio frame structure alternatively.

[0111] Depending on the frame structure and / or the frame configuration in the frame structure, it is possible to perform frequency division duplex (FDD) communication and / or time division duplex (TDD) communication and / or full duplex (FD) communication. FDD communication means that transmissions in different directions (such as uplink and downlink) are performed on different frequency bands. TDD communication means that transmissions in different directions (such as uplink and downlink) are performed in different durations. FD communication means that transmission and reception are performed on the same time-frequency resources, that is, the device can send and receive on the same frequency resources at the same time.

[0112] An example of a frame structure is the frame structure in long-term evolution (LTE), which has the following specifications: the duration of each frame is 10 ms; each frame has 10 subframes, each of which has a duration of 1 ms; each subframe includes two time slots, each of which has a duration of 0.5 ms; each time slot is used to transmit 7 OFDM symbols (assuming normal CP); each OFDM symbol has a symbol duration and a specific bandwidth (or partial bandwidth or bandwidth partition) related to the number of subcarriers and the subcarrier spacing; the frame structure is based on OFDM waveform parameters such as subcarrier spacing and CP length (where CP has a fixed length or finite length option); in TDD, the switching interval between uplink and downlink needs to be an integer multiple of the OFDM symbol duration.

[0113] Another example of a frame structure is the frame structure in new radio (NR), which has the following specifications: multiple subcarrier spacings are supported, each subcarrier spacing corresponds to a corresponding parameter set; the frame structure depends on the parameter set, but in any case, the frame length is set to 10ms, each frame consists of 10 subframes, each subframe is 1ms; the time slot is defined as 14 OFDM symbols; the time slot length depends on the parameter set. For example, the NR frame structure with normal CP 15kHz subcarrier spacing ("parameter set 1") and the NR frame structure with normal CP 30kHz subcarrier spacing ("parameter set 2") are different. For a subcarrier spacing of 15kHz, the time slot length is 1ms; for a subcarrier spacing of 30kHz, the time slot length is 0.5ms. The NR frame structure can have greater flexibility than the LTE frame structure.

[0114] Another example of a frame structure is an exemplary flexible frame structure, such as for 6G networks or higher. In a flexible frame structure, a symbol block may be defined as a minimum duration that may be scheduled in the flexible frame structure. A symbol block may be a transmission unit having an optional redundant portion (e.g., a CP portion) and an information (e.g., data) portion. An OFDM symbol is an example of a symbol block. Alternatively, a symbol block may be referred to as a symbol. Embodiments of a flexible frame structure include different parameters that may be configurable, such as a frame length, a subframe length, a symbol block length, etc. In some embodiments of the flexible frame structure, a non-exhaustive list of possible configurable parameters includes:

[0115] (1) Frame: The frame length does not need to be limited to 10ms. The frame length can be configurable and vary over time. In some embodiments, each frame includes one or more downlink synchronization channels and / or one or more downlink broadcast channels. Each synchronization channel and / or broadcast channel can be transmitted in different directions through different beamforming. The frame length can be more than one possible value and is configured based on the application scenario. For example, an autonomous vehicle may require relatively fast initial access. In this case, the frame length can be set to 5ms for autonomous vehicle applications. For another example, a smart meter on a house may not require fast initial access. In this case, the frame length can be set to 20 for smart meter applications.

[0116] ms.

[0117] (2) Subframe duration: Subframes may or may not be defined in the flexible frame structure, depending on the implementation. For example, a frame may be defined to include time slots but not subframes. In a frame that is defined with subframes, for example, for time domain alignment, the duration of the subframe may be configurable. For example, the length of the subframe may be configured to be 0.1ms or 0.2ms or 0.5ms or 1ms or 2ms or 5ms, etc. In some embodiments, if a subframe is not required in a specific scenario, the subframe length may be defined to be the same as the frame length or may not be defined.

[0118] (3) Time slot configuration: Time slots may or may not be defined in a flexible frame structure, depending on the implementation. In a frame with defined time slots, the definition of the time slots (for example, in terms of duration and / or number of symbol blocks) may be configurable. In one embodiment, the time slot configuration is common to all UEs or a group of UEs. In this case, the time slot configuration information may be sent to the UE in a broadcast channel or a common control channel. In other embodiments, the time slot configuration may be UE-specific, in which case the time slot configuration information may be sent in a UE-specific control channel. In some embodiments, the time slot configuration signaling may be sent together with the frame configuration signaling and / or the subframe configuration signaling. In other embodiments, the time slot configuration may be sent independently of the frame configuration signaling and / or the subframe configuration signaling. In general, the time slot configuration may be system-wide, base station-wide, UE group-wide, or UE-specific.

[0119] (4) Subcarrier spacing (SCS): SCS is a parameter of the scalable parameter set that can allow the SCS to range from 15KHz to 480KHz. The SCS can vary with the spectrum frequency and / or the maximum UE speed to minimize the effects of Doppler shift and phase noise. In some examples, there can be separate transmit frames and receive frames, and the SCS of the symbols in the receive frame structure can be configured independently of the SCS of the symbols in the transmit frame structure. The SCS in the receive frame can be different from the SCS in the transmit frame. In some examples, the SCS of each transmit frame can be half the SCS of each receive frame. If the SCS between the receive frame and the transmit frame is different, the difference does not have to be scaled by a factor of 2, for example, if an inverse discrete Fourier transform (IDFT) is used.

[0120] Instead of fast Fourier transform (FFT) to achieve more flexible symbol duration. Additional examples of frame structures can be used with different SCSs.

[0121] (5) Flexible transmission duration of basic transmission unit: The basic transmission unit can be a symbol block (alternatively referred to as a symbol), which generally includes a redundant portion (referred to as a CP) and an information (e.g., data) portion, although in some embodiments, the CP can be omitted from the symbol block. The CP length can be flexible and configurable. The CP length can be fixed within a frame or flexible within a frame, and the CP length can change with changes in frames, or with changes in frame groups, or with changes in subframes, or with changes in time slots, or dynamically with changes in scheduling. The information (e.g., data) portion can be flexible and configurable. Another possible parameter associated with the symbol block that can be defined is the ratio of the CP duration to the information (e.g., data) duration. In some embodiments, the symbol block length can be adjusted based on channel conditions (e.g., multipath delay, Doppler) and / or delay requirements and / or available duration. For another example, the symbol block length can be adjusted to fit the available duration in a frame.

[0122] (6) Flexible switching gap: A frame may include a downlink part for downlink transmission sent by a base station and an uplink part for uplink transmission sent by a UE. There may be a gap between each uplink part and the downlink part, which is called a switching gap. The switching gap length (duration) may be configurable. The switching gap duration may be fixed within a frame or flexible within a frame. The switching gap duration may change with changes in frames, changes in frame groups, changes in subframes, changes in time slots, or dynamically changes with changes in scheduling.

[0123] Cell / Carrier / Bandwidth part (BWP) / Occupied bandwidth

[0124] Devices such as base stations can provide cell coverage. Wireless communication with a device can be performed on one or more carrier frequencies. A carrier frequency is called a carrier. Alternatively, a carrier can be called a component carrier (CC). A carrier can be characterized by its bandwidth and a reference frequency, such as the center frequency, the lowest frequency, or the highest frequency of the carrier. The carrier can be on a licensed spectrum or an unlicensed spectrum. Additionally or alternatively, wireless communication with a device can be performed on one or more bandwidth parts (BWP). For example, a carrier can have one or more BWPs. In general, wireless communication with a device can be performed on a spectrum. The spectrum can include one or more carriers and / or one or more BWPs.

[0125] The cell may include one or more downlink resources and optionally one or more uplink resources, or the cell may include one or more uplink resources and optionally one or more downlink resources, or the cell may include both one or more downlink resources and one or more uplink resources. For example, the cell may include only one downlink carrier / BWP, or only one uplink carrier / BWP, or multiple downlink carriers / BWP, or multiple uplink carriers / BWP, or one downlink carrier / BWP and one uplink carrier / BWP, or one downlink carrier / BWP and multiple uplink carriers / BWP, or multiple downlink carriers / BWP and one uplink carrier / BWP, or multiple downlink carriers / BWP and multiple uplink carriers / BWP. In some embodiments, alternatively or additionally, the cell may include one or more sidelink resources, and the one or more sidelink resources include sidelink transmission and reception resources.

[0126] A BWP is a set of continuous or non-contiguous frequency subcarriers on a carrier, or a set of continuous or non-contiguous frequency subcarriers on multiple carriers, or a set of non-contiguous or continuous frequency subcarriers, and the BWP may have one or more carriers.

[0127] In some embodiments, a carrier may have one or more BWPs, for example, a carrier may have a bandwidth of 20 MHz and consist of one BWP, or a carrier may have a bandwidth of 80 MHz and consist of two adjacent continuous BWPs, etc. In other embodiments, a BWP may have one or more carriers, for example, a BWP may have a bandwidth of 40 MHz and consist of two adjacent continuous carriers, each of which has a bandwidth of 20 MHz. In some embodiments, a BWP may include non-continuous spectrum resources consisting of non-continuous multi-carriers, wherein the first carrier in the non-continuous multi-carrier may be in the mmW band, the second carrier may be in the low frequency band (e.g., the 2 GHz band), the third carrier (if present) may be in the THz band, and the fourth carrier (if present) may be in the visible light band. The resources in a carrier belonging to a BWP may be continuous or non-continuous. In some embodiments, a BWP has non-continuous spectrum resources on a carrier.

[0128] Wireless communication can be performed over an occupied bandwidth. The occupied bandwidth can be defined as the width of the frequency band such that below a lower frequency limit and above an upper frequency limit, the average power transmitted is equal to a specified percentage □ / 2 of the total average transmitted power, such as □ / 2 being 0.5%.

[0129] The carrier, BWP, or occupied bandwidth may be dynamically signaled by a network device (e.g., a base station) (e.g., in physical layer control signaling) (e.g., downlink control information (DCI)), or semi-statically signaled (e.g., in radio resource control (RRC) signaling, or in the medium access control (MAC) layer), or predefined based on an application scenario; or determined by the UE as a function of other parameters known to the UE, or may be fixed by, for example, a standard.

[0130] Artificial intelligence (AI) and / or machine learning (ML)

[0131] The number of new devices in future wireless networks is expected to grow exponentially, and the functionality of devices is expected to become increasingly diverse. In addition, many new applications and use cases are expected to emerge, with more diverse quality of service requirements compared to 5G applications / use cases. These will generate new key performance indications (KPIs) for future wireless networks (e.g., 6G networks), which may be extremely challenging. AI technologies, such as ML technologies (e.g., deep learning), have been introduced into telecommunication applications with the goal of improving system performance and efficiency.

[0132] In addition, antenna and bandwidth capabilities continue to advance, making it possible to carry out more and / or better communications over wireless links. In addition, advances in computer architecture and computing power continue, for example, the introduction of general-purpose graphics processing units (GP-GPUs). Future generations of communication devices may have more computing and / or communication capabilities than previous generations, which may allow the use of AI to implement air interface components. Future generations of networks may also obtain more accurate and / or new information (compared to previous generations of networks), which can become the input basis of AI models, such as the physical speed (speed / velocity) at which the device moves, the link budget of the device, the channel conditions of the device, one or more device capabilities and / or the type of services that need to be supported, sensing information and / or positioning information, etc. In order to obtain sensing information, the TRP can send a signal to the target object (e.g., a potential UE), and the TRP or another network device calculates the angle (for beamforming of the device), the distance between the device and the TRP, and / or Doppler frequency deviation information based on the reflection of the signal. Positioning information is sometimes referred to as positioning, and it can be obtained in a variety of ways, such as positioning reports from the UE (e.g., reports of the UE's GPS coordinates), using a positioning reference signal (PRS), using the above-mentioned sensing, tracking and / or predicting the location of the device, etc.

[0133] AI technology (including ML technology) can be applied to communications, including AI-based communications in the physical layer and / or AI-based communications in the MAC layer. For the physical layer, AI communication can be intended to optimize component design and / or improve algorithm performance. For example, AI can be used to implement the following aspects: channel coding, channel modeling, channel estimation, channel decoding, modulation, demodulation, MIMO, waveforms, multiple access, physical layer element parameter optimization and update, beamforming, tracking, sensing and / or positioning, etc. For the MAC layer, AI communication can be intended to use AI capabilities to learn, predict and / or make decisions to use better strategies and / or optimal solutions to solve complex optimization problems, such as optimizing functions in the MAC layer. For example, AI can be applied to implement: intelligent TRP management, intelligent beam management, intelligent channel resource allocation, intelligent power control, intelligent spectrum utilization, intelligent MCS, intelligent HARQ strategy and / or intelligent transmit / receive mode adaptation, etc.

[0134] In some embodiments, the AI ​​architecture may involve multiple nodes, where the multiple nodes may be organized in one of two modes: centralized and distributed, both of which may be deployed in an access network, core network, or edge computing system or a third-party network. Centralized training and computing architectures are subject to potentially large communication overhead and strict user data privacy restrictions. Distributed training and computing architectures may include several frameworks, such as distributed machine learning and federated learning. In some embodiments, the AI ​​architecture may include an intelligent controller that may execute as a single agent or multiple agents based on joint optimization or individual optimization. New protocols and signaling mechanisms are required so that the corresponding interface links can be personalized through custom parameters to meet specific needs, while minimizing signaling overhead and maximizing the overall system spectrum efficiency through personalized AI technology.

[0135] In some embodiments herein, new protocols and signaling mechanisms are provided for operating within and switching between different operating modes for AI training, including switching between training and normal operating modes, and for measurement and feedback to accommodate different possible measurements and information that may require feedback, depending on the implementation.

[0136] AI Training

[0137] Reference again Figure 1 and 2 , the embodiments of the present disclosure may be used to implement AI training involving two or more communication devices in the communication system 100. For example, Figure 5Four EDs are shown in communication with a network device 452 in a communication system 100 according to one embodiment. The four EDs are shown as respective different UEs and will be referred to hereinafter as UEs 402, 404, 406, and 408. However, an ED does not necessarily have to be a UE.

[0138] Network device 452 is part of a network (e.g., wireless access network 120). Network device 452 can be deployed in an access network, a core network, an edge computing system, or a third-party network, depending on the implementation. Network device 452 can be a T-TRP or a server (or a part thereof). In one example, network device 452 can be a T-TRP 170 or a NT-TRP 172 (or implemented therein). In another example, network device 452 can be a T-TRP controller and / or a NT-TRP controller that can manage T-TRP 170 or NT-TRP 172. In some embodiments, the components of network device 452 can be distributed. UE 402, 404, 406, and 408 can communicate directly with network device 452, for example, if network device 452 is part of a T-TRP serving UE 402, 404, 406, and 408. Alternatively, UEs 402, 404, 406, and 408 may communicate with network device 352 through one or more intermediate components (e.g., through T-TRP and / or through NT-TRP, etc.). For example, network device 452 may send and / or receive information (e.g., control signaling, data, training sequences, etc.) to / from one or more of UEs 402, 404, 406, and 408 through a backhaul link and a wireless channel interposed between network device 452 and UEs 402, 404, 406, and 408.

[0139] As described above, each UE 402, 404, 406 and 408 includes a corresponding processor 210, a memory 208, a transmitter 201, a receiver 203 and one or more antennas 204 (or alternatively, a panel). For simplicity, only the processor 210, the memory 208, the transmitter 201, the receiver 203 and the antenna 204 of the UE 402 are shown, but the other UEs 404, 406 and 408 also include the same corresponding components.

[0140] For each UE 402, 404, 406 and 408, the communication link between the UE and the corresponding TRP in the network is an air interface. The air interface typically includes many components and associated parameters that together specify how to send and / or receive transmissions over a wireless medium.

[0141] Figure 5The processor 210 of the UE in the embodiment implements one or more air interface components on the UE side. The air interface component configures and / or implements transmission and / or reception on the air interface. Examples of air interface components are described herein. The air interface component may be located in the physical layer, for example, a channel encoder (or decoder) that implements the air interface coding component of the UE, and / or a modulator (or demodulator) that implements the air interface modulation component of the UE, and / or a waveform generator that implements the air interface waveform component of the UE, etc. The air interface component may be in a high layer such as a MAC layer or be part of it, for example, a module that implements channel prediction / tracking, and / or a module that implements a retransmission protocol (for example, an air interface HARQ protocol component that implements the UE), etc. The processor 210 also directly executes (or controls the UE to execute) the UE-side operations described herein.

[0142] The network device 452 includes a processor 454, a memory 456, and an input / output device 458. The processor 454 implements or instructs other network devices (e.g., T-TRP) to implement one or more air interface components on the network side. Compared with another UE, the air interface component implementation mode of one UE on the network side may be different. The processor 454 directly executes (or controls the network component to execute) the network side operations described herein.

[0143] Processor 454 may be implemented by the same or different one or more processors for executing instructions stored in memory (e.g., memory 456). Alternatively, some or all of processor 454 may be implemented using dedicated circuits such as programmed FPGAs, GPUs, or ASICs. Memory 456 may be implemented by volatile and / or non-volatile storage elements. Any suitable type of memory may be used, such as RAM, ROM, hard disk, optical disk, on-processor cache, etc.

[0144] The input / output devices 458 allow interaction with other devices by receiving (input) and sending (output) information. In some embodiments, the input / output devices 458 can be implemented by a transmitter and / or a receiver (or transceiver) and / or one or more interfaces (e.g., a wired interface to an internal network or to the Internet, etc.). In some implementations, the input / output devices 458 can be implemented by a network interface, which may be implemented as a network interface card (NIC) and / or a computer port (e.g., a physical socket to which a plug or cable is connected) and / or a network socket, etc., depending on the implementation.

[0145] The network device 452 and the UE 402 have the capability to implement one or more AI-enabled processes. Figure 5In the embodiment of the present invention, the network device 452 and the UE 402 include ML modules 510 and 500 respectively. The ML module 510 is implemented by the processor 210 of the UE 402, and the ML module 500 is implemented by the processor 454 of the network device 452. Figure 5 , ML module 510 is shown within processor 210, and ML module 500 is shown within processor 454. For example, ML modules 510 and 500 execute one or more AI / ML algorithms to perform one or more AI-enabled processes, such as AI-enabled link adaptation, to optimize the communication link between the network and UE 402.

[0146] ML modules 510 and 500 can be implemented using AI models. The term AI model can refer to a computer algorithm for receiving defined input data and outputting defined reasoning data, wherein the parameters of the algorithm (e.g., weights) can be updated and optimized by training (e.g., using a training data set, or using data collected in real life). The AI ​​model can be implemented using one or more neural networks (e.g., including deep neural networks (DNNs), recurrent neural networks (RNNs), convolutional neural networks (CNNs), and combinations thereof) and using various neural network architectures (e.g., autoencoders, generative adversarial networks, etc.). Various techniques can be used to train the AI ​​model in order to update and optimize its parameters. For example, back propagation is a common technique for training DNNs, in which a loss function is calculated between the reasoning data generated by the DNN and a certain target output (e.g., ground truth data). The gradient of the loss function is calculated relative to the parameters of the DNN, and the parameters are updated using the calculated gradient (e.g., using a stochastic gradient descent (SGD) algorithm) to minimize the loss function.

[0147] In some embodiments, the AI ​​model covers a neural network for machine learning. A neural network consists of multiple computing units (which may also be referred to as neurons) arranged in one or more layers. The process of receiving input at an input layer and generating output at an output layer can be referred to as forward propagation. In forward propagation, each layer receives an input (which may have any suitable data format, such as a vector, matrix, or multidimensional array) and performs calculations to generate an output (which may have a different dimension than the input). The calculations performed by the layer typically involve applying (e.g., multiplying) a set of weights (also called coefficients) to the input. Except for the first layer (i.e., input layer) of the neural network, the input of each layer is the output of the previous layer. The neural network may include one or more layers between the first layer (i.e., input layer) and the last layer (i.e., output layer), which may be referred to as inner layers or hidden layers. For example, Fig. 6A An example of a neural network 600 including an input layer, an output layer, and two hidden layers is depicted. In this example, it can be seen that the output of each of the three neurons in the input layer of the neural network 600 is included in the input vector of each of the three neurons in the first hidden layer. Similarly, the output of each of the three neurons in the first hidden layer is included in the input vector of each of the three neurons in the second hidden layer, and the output of each of the three neurons in the second hidden layer is included in the input vector of each of the two neurons in the output layer. As described above, the basic computational unit in a neural network is a neuron, such as Fig. 6A As shown in 650. Figure 6B An example of a neuron 650 that can be used as a building block of the neural network 600 is shown. Figure 6B As shown in , in this example, neuron 650 takes vector x as input and performs a dot product with the associated weight vector w. The final output y of the neuron is the result of the activation function f() on the dot product. Various neural networks can be designed to have various architectures (e.g., various numbers of layers, each layer performing various functions).

[0148] For example, an autoencoder (AE) is a neural network with a specific architecture that is suitable for a specific application. Unlike neural networks for classification or regression, in most traditional use cases, the training goal of an AE is to output a vector with maximum accuracy. Reproduce its input vector x at . Note that the AE has one hidden layer, called the latent space z, whose dimensionality is smaller than the dimensionality of the input layer. The latent space can be thought of as a compressed representation, and the layers before and after the latent space are the encoder and decoder, respectively. In many cases, it is desirable to minimize the size of the latent space or dimensionality while maintaining the accuracy of the decoder. Figure 7An example of an AE 700 is shown that includes an encoder 702, a latent space 704, and a decoder 706. In this example, the dimensionality of the encoder 702 input is 5, which is reduced to 3 at the latent space 704 and expanded to 5 again by the decoder 706.

[0149] Neural networks are trained to optimize the parameters (e.g., weights) of the neural network. This optimization is performed in an automated manner and can be referred to as machine learning. The training of a neural network includes forward propagating input data samples to generate output values ​​(also called predicted output values ​​or inference output values), and comparing the generated output values ​​with known or expected target values ​​(e.g., ground truth values). A loss function is defined to quantitatively represent the difference between the generated output value and the target value, and the goal of training the neural network is to minimize the loss function. Backpropagation is an algorithm used to train a neural network. Backpropagation is used to adjust (also called update) the values ​​of parameters (e.g., weights) in a neural network so that the calculated loss function becomes smaller. Backpropagation involves calculating the gradient of the loss function relative to the parameters to be optimized, and updating the parameters using a gradient algorithm (e.g., gradient descent) to reduce the loss function. Backpropagation is performed iteratively so that the loss function converges or is minimized over multiple iterations. After the training conditions are met (e.g., the loss function has converged, or a predefined number of training iterations have been performed), the neural network is considered trained. The trained neural network can be deployed (or executed) to generate inference output data from the input data. In some embodiments, training of the neural network may continue even after the neural network has been deployed, such that parameters of the neural network may be repeatedly updated with the latest training data.

[0150] Reference again Figure 5 In some embodiments, the UE 402 and the network device 452 may exchange information for training purposes. The information exchanged between the UE 402 and the network device 452 is implementation specific, and it may not have a human-understandable meaning (e.g., it may be intermediate data generated during the execution of the ML algorithm). Additionally or alternatively, the information exchanged may not be predefined by the standard, for example, bits may be exchanged, but the bits may not be associated with a predefined meaning. In some embodiments, the network device 452 may provide or indicate one or more parameters to the UE 402 to be used in the ML module 510 implemented at the UE 402. As an example, the network device 452 may send or indicate updated neural network weights to be implemented on the UE side in a neural network executed by the ML module 510 in order to attempt to optimize one or more aspects of modulation and / or coding for communications between the UE 402 and the T-TRP or NT-TRP.

[0151] In some embodiments, UE 402 may implement AI itself, such as performing learning, while in other embodiments, UE 402 may not be able to perform learning itself, but may be able to operate in conjunction with AI implementation on the network side, for example, by receiving from the network a configuration of an AI model (e.g., a neural network or other ML algorithm) implemented by ML module 510, and / or by assisting other devices (e.g., network devices or other UEs with AI capabilities) in training AI models (e.g., neural networks or other ML algorithms) by providing requested measurements or observations. For example, in some embodiments, UE 402 may not be able to implement learning or training itself, but UE 402 may receive trained configuration information of an ML model determined by network device 452 and execute the model.

[0152] Although Figure 5 The examples in assume AI / ML capabilities on the network side, but it is possible that the network itself does not perform training / learning, but rather the UE can perform learning / training itself, possibly using dedicated training signals sent from the network. In other embodiments, end-to-end (E2E) learning can be implemented by the UE and the network device 452.

[0153] Scheduling

[0154] As the number and density of wireless communication devices increase, supporting high-quality wireless communications using traditional wireless systems becomes increasingly challenging.

[0155] Machine-to-machine (M2M) communication can be a type of high-density wireless communication. M2M communication is a technology that implements a network for collecting information from devices (e.g., sensors, smart meters, Internet of Things (IoT) devices, and / or other low-end devices) that are typically deployed on a large scale and densely, and for sending the information captured by these devices to other applications in the network. M2M networks can be wired or wireless and may have a relatively large geographic distribution (e.g., across a country or across the world). M2M communication typically does not involve direct human intervention to collect information.

[0156] 5G New Radio (NR) systems include capabilities to support massive machine type communications (mMTC), which connect large numbers (e.g., millions or billions) of IoT devices through wireless systems. It is expected that in the near future, the amount of M2M communications conducted over the air will exceed the amount of communications associated with humans. For example, 6G systems are expected to connect more IoT devices than mobile phones. In 6G, high-density IoT deployments are expected to spawn many innovative applications, profoundly reshaping many industries and societies. Some forecasts predict that deployment densities in 6G systems could reach 1 km per 1000 Mbps. 2 10 9 It will be a challenge for 6G systems to support such high-density IoT deployments, where thousands or tens of thousands of IoT devices may simultaneously send their data back to the network over a shared wireless channel.

[0157] Using AI, for example, by implementing an AI model as described above, various processes such as transmission scheduling can be AI-enabled. Some examples of possible AI / ML training processes and over-the-air information exchange processes between devices according to embodiments of the present disclosure are described below to facilitate AI-enabled scheduling of a large number of densely deployed IoT devices.

[0158] Proportionality-Fairness

[0159] As previously referenced Figure 3 As discussed, the network scheduler, e.g. Figure 3 The scheduler 253 shown is usually responsible for allocating wireless resources among all its associated terminals or devices. In the past 20 years, the scheduling algorithm based on proportional-fairness (PF) has become the industry standard, and most currently deployed schedulers use the scheduling algorithm based on PF or a derivative algorithm thereof.

[0160] PF-based schedulers aim to maintain absolute fairness among all devices by allocating a portion of the radio resources to devices that is proportional to their relative requests. For example, suppose 100 active devices request radio resources at the same time. If device A requests 5% of the total, a PF-based scheduler will typically allocate 5% of the radio resources to the device. Classical information theory proves that PF is a maximum-likelihood (ML) criterion optimization, which is theoretically optimal and one of the reasons why the PF algorithm has been the benchmark scheduling algorithm for many years.

[0161] However, high-density IoT deployments and the new applications potentially associated with such deployments will challenge the PF algorithm by challenging the assumption that requests from each device are independent and equally important. This assumption is usually correct for mobile devices that can generally be considered independent, such as smartphones or tablets. The PF algorithm produces system gain by enforcing proportional fairness among mobile devices, prohibiting the network from discriminating against any mobile device. To illustrate this, the simultaneous requests from multiple mobile devices are considered as a multi-dimensional request distribution, and the entropy of the distribution is measured. According to information theory, the more heterogeneous the distribution, the lower the entropy of the distribution, and the more system gain margin the PF scheduler can achieve.

[0162] For example, Figure 8 The entropy of the 2-dimensional (2-device) request distribution of two devices, device A and device B, is shown. In the case of heterogeneous distribution of requests between device A and device B, such as 80% and 20% (for example, device A is playing a video and device B is sending a message), the PF-based scheduler can bring significant system gains, such as Figure 8 In contrast, in the case of roughly homogeneous distributions of requests, such as 45% and 55% (e.g., if both devices are downloading music), the PF-based scheduler can achieve only small system gains.

[0163] High Entropy IoT Request Distribution

[0164] Next, we consider a typical multi-request distribution for IoT deployments. In some typical IoT scenarios, many IoT devices will measure or observe a common object, phenomenon, environment, goal, or event from different perspectives, angles, formats, or types. Since each device considers itself as informative as any other device, the resulting entropy of the multi-request distribution is likely to be too high, making it impossible for a PF-based scheduler to provide any significant system gain.

[0165] Informational

[0166] Unlike mobile devices, IoT devices are usually not equally informative to downstream tasks. For example, when measuring or observing a common object, some more advantageously located devices may be more informative than other devices for downstream tasks based on object information. Similarly, some more disadvantageously located devices may be completely irrelevant to downstream tasks, while some similarly located devices may be redundant with each other. Ideally, when scheduling devices to obtain data for downstream tasks, the scheduler would avoid scheduling irrelevant and redundant devices, and prioritize scheduling devices that can provide informative data for downstream tasks. However, this is not achievable with traditional PF schedulers.

[0167] For example, Fig. 9An exemplary scenario is shown in which a plurality of devices 900 are deployed and can potentially be scheduled to send information about measurements or observations of objects or phenomena within an observation area 902. In this example, the downstream task can be related to detecting objects in the observation area 902. Fig. 9 As shown, devices B, C, and D are informative for the task of detecting manifold "O" 904 in observation area 902. However, it can also be seen that device A is irrelevant to the task, while devices B and C are redundant with each other for the purpose of the task. If network device 452 were to utilize a PF-based scheduler to schedule uplink transmissions from device 900, the network device would evenly allocate the same radio resources among devices A, B, C, and D.

[0168] Contribution to path loss

[0169] The environment of a network device (terrain, hills, buildings, woods, etc.) may cause extensive path loss among its associated IoT devices. For example, devices with high information content may suffer more severe path loss than devices with low information content. In such a scenario, despite the high information content, the device will be of no use if the network cannot successfully receive the device's uplink signal. Therefore, it is important to remember that in wireless communications, being informative does not necessarily mean being contributing. In other words, although the information collected by the deployed device may be informative, if the device cannot successfully transmit the information to the network, the device will not contribute to the downstream tasks that rely on such information. In such a scenario, it may be more valuable to schedule a device with low information content but low path loss than to schedule a device with very high information content but high path loss.

[0170] For example, Fig.10 1000 in the surroundings of network device 452. Fig. 9 1000. An example of the potential impact of different channel conditions in an IoT deployment scenario. Specifically, in this scenario, a device deployed between network device 452 and hill 1000 may be in region 1002 where it experiences very low path loss but does not provide information to the downstream detection task. On the other hand, although device B is informative, it may not contribute due to the depth 1004 behind hill 1000. In this scenario, rather than scheduling device B, which is informative but does not contribute, it is better to schedule device C, which is redundant and has less information.

[0171] Contribution to the task

[0172] Aspects of the present disclosure introduce the concept of contribution metrics, which are metrics used to compare two IoT devices. Classical information theory makes it difficult to evaluate, score, or compare two sources unless they are strictly in the same context. As a simple example, we can compare two English sentences and even score the differences (gaps) between them because the two sentences use the same vocabulary and grammar (context). However, we may not be able to compare an English sentence and a Chinese sentence, even if they express the same semantics, because English and Chinese have very different vocabularies and grammars (different contexts).

[0173] The contribution depends on a given downstream task; that is, how much the information provided by the source contributes to the downstream processing of the entire fused information to complete a specific task. For example, imagine a scenario where thousands of IoT devices are monitoring a vast forest and providing information based on their monitoring to the network. Their measurements may eventually be fused into the network, where some applications will process these measurements to complete some tasks. For example, some tasks may classify the current forest state, some tasks may reconstruct a real-time forest model in a virtual world, and some tasks may raise an alarm when a forest fire is detected. Different tasks form their own context to score the contribution of any source fused into the network. For example, a thermometer IoT device that is considered to contribute to the forest fire alarm task may not be of much use for the forest wildlife population size task. Therefore, new tasks may require new scheduling algorithms specific to scheduling devices for said tasks.

[0174] For example, refer to Fig.11A , it can be seen that the task associated with classifying the "O" manifold 904 may require information from devices B, E, F, and D. On the other hand, reference Fig. 11B , it can be seen that the task of detecting the presence of manifold 904 may only require devices E and F.

[0175] Incomplete data

[0176] Data-driven approaches can potentially be used to identify contributing devices for downstream tasks with channels with varying path losses. Ideally, the goal of such an approach is to accumulate a sufficient amount of raw data from all IoT devices over a period of time to be able to determine which devices contribute and which do not.

[0177] However, the extremely high density and large number of IoT devices make it almost impossible for wireless systems to accumulate complete data sets. This is partly due to the high cost of wireless resources and partly due to random adverse channels. Data packets from devices that suffer from severe path loss may be lost or fail to arrive at the network on time.

[0178] For example, Fig.12 Shows Fig.10 An example of an IoT deployment in which devices G, H, I, and J are currently experiencing high erasure channel conditions and are therefore generally unable to successfully send data to network device 452. In this scenario, if network device 452 randomly allocates uplink radio resources among a certain percentage (e.g., 80%) of candidate devices within a given sampling / measurement interval, some devices, such as devices G, H, I, and J that are currently experiencing high packet loss rates, may not be able to successfully send their information back to the network even if they are allocated radio resources. Additionally or alternatively, some devices may not be ready to send measurement information when they are allocated radio resources (e.g., due to measurement and / or processing limitations).

[0179] Therefore, despite the incomplete dataset, a robust scheduling algorithm should be able to identify the devices that contribute most to the downstream tasks.

[0180] PF-based scheduling is not suitable for IoT deployment

[0181] The highly dense deployment of a large number of IoT devices expected in 6G networks will bring several new challenges to wireless network schedulers. If these deployments adopt a PF-based scheduler, since each device has little knowledge of its contribution to the downstream tasks, each device considers itself to be as contributing as any other device. Therefore, the resulting high-entropy multi-request distribution from the devices will cause the PF-based scheduler to evenly distribute the available wireless resources into small fragments that it will allocate among all requesting devices. Finally, as mentioned above, Figure 8 As discussed, the system gain will be reduced and the system will become less efficient.

[0182] Sparsity of phenomenon

[0183] A fundamental fact is that only a few laws govern a rich physical environment, that is, sparse laws produce rich data. Sparsity is hidden under the large amount of data of the object. General learning is to associate a large amount of data to the task; while effective learning is to associate data sparsity to the task.

[0184] When faced with a large number of IoT devices, different path losses on the channel, and a given task, it is desirable to have a scheduler that can learn some of the most contributing devices (or sensors) that can successfully provide enough information about the observed object to the processor to complete the task. If possible, the scheduler will not have to allocate wireless resources to all of its associated IoT devices, but can instead allocate wireless resources to a few devices that it has identified as contributing to the task. This type of selective scheduling can significantly reduce or alleviate system overload.

[0185] Contribution-based scheduling

[0186] Aspects of the present disclosure address several scheduling challenges anticipated in highly dense IoT deployments by providing methods to learn which devices contribute to a given downstream task and schedule based on their contribution.

[0187] Consider the following IoT scenario:

[0188] -A large number of IoT devices, terminals or sensors (used interchangeably in this disclosure) are connected to the network through wireless channels.

[0189] -These IoT devices are monitoring, measuring, observing and collecting information about a common object, environment, goal or phenomenon (used interchangeably in this disclosure).

[0190] - In order to save radio resources, the network will schedule (allocate radio resources for the transmission of these measurements) as few IoT devices as possible for a given task. Thus, the network's scheduler will allocate the smallest radio resources to the most contributing IoT devices to send their measurements to the network, which will complete the task with these sent measurements.

[0191] - The path loss, channel conditions, erasure rate or packet loss rate (used interchangeably in this disclosure) from the scheduled IoT devices to the network is not uniformly (differently) random, i.e. some devices suffer more path loss (more unfavorable channel conditions) than others.

[0192] -The information received from the scheduled IoT devices will be fused into the network's decoder or processor (used interchangeably in this disclosure), which will accomplish specific downstream tasks, purposes or goals (used interchangeably in this disclosure).

[0193] -Each IoT device, whether scheduled or not, has little idea of ​​how much it contributes to downstream tasks. Therefore, each device continues to request radio resources to make its own measurements.

[0194] - The scheduler of the network device sends a scheduling message or signaling (used interchangeably in this disclosure) to the scheduled IoT devices, and it can generally be assumed that there is no packet loss on the downlink channel from the network to these IoT devices. Advantages provided by various aspects of the present disclosure and / or problems that various aspects of the present disclosure can solve or alleviate include:

[0195] -A scheduler that allocates minimal resources to a small number of IoT devices whose measurements contribute most to downstream tasks.

[0196] -When selecting those devices that contribute most to the downstream tasks, the scheduler can take into account both their informativeness and their individual channel conditions. The informativeness of the measurements and the packet loss rate of the channel from the device to the network (uplink) are optimized. (Contribution = informativeness + packet loss rate)

[0197] -Learning device selection using incomplete datasets of measurement results.

[0198] -When selecting devices to schedule, the scheduler can also score and group devices based on their contribution. Thus, the scheduler can choose to schedule the group of devices with the largest contribution in the first batch. If the information received from this group is not sufficient to complete the task, the scheduler can proceed to schedule a second group of devices that are rated as contributing less than the devices in the first largest contributing group. In this scenario, the information received from the two groups can then be combined to complete the downstream task in different random channel distributions.

[0199] Although scheduling the device with the largest contribution looks like sampling, it is more challenging and complex than sampling. In sampling problems, the most representative samples are usually kept for reconstruction. In scheduling problems, the scheduler may need to complete arbitrary downstream tasks other than simple reconstruction tasks, preferably while using the least amount of radio resources, which means two inseparable functions: the scheduler that selects the device with the largest contribution and the decoder that fuses and processes the received data for a given task.

[0200] In some embodiments of the present disclosure, a deep neural network (DNN) or an autoencoder (AE) is used to train: a scheduler that selects the IoT devices with the greatest contribution from all candidate IoT devices and scores these IoT devices; and a decoder that fuses and processes the information received from the selected (scheduled) devices. The scheduler is based on the first layer of the trained DNN, and the decoder is implemented by the subsequent layers of the DNN. The downstream task is defined as the training objective of this DNN. It can be supervised with labeled data or unsupervised without labeled data. The input training data consists of multiple samples. Due to limited wireless resources and erasure channels, each sample may contain less than the original measurement results of all candidate devices.

[0201] In some aspects of the present disclosure, the selector and decoder are trained together in one DNN by a stochastic gradient descent (SGD) training algorithm. Backpropagation backpropagates the final training target (task target) to each neuron from the last layer to the first layer. As discussed in further detail below, the first layer corresponds to the scheduler / selector, and the remaining layers of the DNN correspond to the decoder. Accordingly, the first layer is a selector, which can be called a scheduling layer, which selects a subset of devices from a set of N candidate devices and sends information from the selected devices to the decoder. Each candidate device can be regarded as an input dimension of the scheduling layer (selector), and only data from the selected device can be output from this layer.

[0202] Fig.13 An example of a DNN configured as an autoencoder 1100 is shown, the DNN including a first scheduling layer serving as a scheduler 1102 and multiple decoding layers serving as fusers and processors 1104 for information received from devices selected by the scheduling layer.

[0203] Unlike a convolutional layer or a max pooling layer that compresses the dimensions from its input to its output by nonlinearly composing or combining several input dimensions into one output dimension (via a nonlinear activation function), the scheduling layer implementing scheduler 1102 is linear and fully connected, so that each of its K output dimensions is a linear combination of all its N input dimensions. For example, Fig.13 As shown, each of the K output dimensions of the scheduling layer (i.e., y j , j = 1, 2, ... K) are N inputs x1, x2, ... N The linear combination of these inputs is represented by their N corresponding weights w 1,j 、w 2,j ...w N,j Weighted so that:

[0204]

[0205] However, as discussed in further detail below, training of the autoencoder 1100 can be performed in a manner that forces the N weights or coefficients of the linear combination of each output dimension to be polarized so that only one input is retained and the rest are eliminated; that is, inputs with weights approaching 1 survive and the rest with weights approaching 0 die.

[0206] In some embodiments, the sparsity of the selection can be enhanced by a training hyperparameter, referred to herein as temperature (or other names for similar controls), to enhance convergence to local minima (polarization). For example, over training, the training temperature can be gradually cooled so that only 1 of the N weights of the linear combination of the output dimensions survives at the end of training. Fig.14 An example of epoch-by-epoch cooling during training is shown in FIG. Fig.14 As shown, during the initial training epochs, all devices may be active, and as the temperature parameter cools epoch by epoch, the number of active devices decreases as less contributing devices are excluded or deselected.

[0207] In practice, although many different cooling strategies can be adopted, an exponential decay strategy is used in the example discussed below. After training is completed, the first layer ends with a selector that selects K devices from the N candidate devices (although a given device may be selected more than once among the K selected devices, as discussed in further detail below), and the rest of the layer acts as a corresponding decoder that fuses and processes the information sent from the selected devices. Based on the trained selector, the scheduler will allocate uplink radio resources and send scheduling messages or signaling to the selected devices.

[0208] In preparing training and test datasets for this deep neural network, the network can collect raw measurements from a set of N candidate devices over several time intervals. However, in practice it is difficult to obtain complete training and test datasets, i.e., every raw measurement from all N devices, partly due to the high communication overhead required to do so and its associated costs, and partly due to random packet loss on different channels. In practice, for example, the network can allocate radio resources that allow a maximum of 70% raw measurements from N devices in each time interval. However, the network may collect less than 70% of the devices, partly due to the random and different path loss distributions of unfavorable channels. Packets for some devices may be more likely to be lost than others. Logically, different packet loss rate distributions will be embedded in the training and test datasets. Since the deep neural network uses this training dataset for learning, the selectors for K devices out of N devices will reflect different packet loss rate distributions.

[0209] Usually, the number of devices K required is task-dependent, that is, the number of devices selected depends on the downstream task. Intuitively, since classification tasks are usually easier than reconstruction tasks, and detection tasks are easier than classification tasks, K is usually 重构 >K 分类 >K 检测is correct. In fact, this solution can be left to the deep neural network itself. At the scheduling (selector) layer, it is not forbidden to connect one input dimension to more than one output dimension. After training, it is very likely that more than two output dimensions come from the same input dimension (device). In other words, some devices are selected more than once. Given a specific task (reflected in the training objective of the deep neural network), it is possible to determine the minimum number of devices K that will not trigger any repeated selections. min This K min The value is considered to be the minimum number of scheduled devices to complete a specific task in the training dataset (embedded by different distributions of unfavorable channels). Different tasks usually have different K min .

[0210] Fig.15 It shows that when the output dimension K of the scheduling layer is larger than K of the given task min Examples of duplicate selection of contributing devices that may occur when Fig.15 As shown in Figure 2, when the output dimension K is equal to Kmin, training causes the scheduling layer to select 4 non-repeating devices A, B, C, and D. However, if the output dimension K increases to K1, where K1>K min , then training again results in the selection of devices A, B, C, and D, but devices A and C are both selected multiple times. Similarly, if the output dimension K is further increased to K2, where K2>K1, then devices A, B, C, and D are again selected, but in this case, devices A, B, and C are all selected multiple times.

[0211] like Fig.15 As shown, when the network deliberately sets K>>K min When training a deep neural network, some devices may be selected repeatedly. The more times a device is selected, the greater its contribution. Therefore, the network can score and cluster the selected devices based on the number of times they are selected for a given task.

[0212] Fig.16 It shows that the output dimension K of the scheduling layer is much larger than K min An example of a scenario and a ranking of contributing devices that can be done based on the number of times the scheduling layer selects each contributing device. In this example, the scheduling layer selects 4 devices A, B, C, and D, but device A is selected 4 times, devices B and C are both selected 2 times, and device D is only selected 1 time.

[0213] Training can be divided into several groups of devices based on the contribution of the device to the task (number of selections). For example, based on the selection of 4 output dimensions by the scheduling layer, device A may be rated as contributing more than devices B and C, which are both selected by 2 output dimensions. Similarly, device D, which is selected by only 1 output dimension, may be rated as contributing less than devices B and C. The scheduler that allocates wireless resources among all selected surviving devices can allocate them proportionally based on their contribution rather than their requests. For example, the scheduler can allocate more bandwidth (lower coding rate, or higher transmission power) to a group including devices selected by 4 output dimensions (e.g., device A) than to a lower-ranked group including devices selected by 2 output dimensions (e.g., device B and device C).

[0214] If the network can cluster the selected surviving devices into groups based on their contribution to the task, the scheduler can take advantage of the opportunity to schedule the group of devices with the largest contribution in the first group (K1). If the information sent from the first batch of devices provides enough confidence for the decoder, there is no need to schedule any other devices; otherwise, the scheduler can continue to schedule the secondary group (K2) of devices with smaller contributions.

[0215] Fig.17 An example is shown where contributing devices have been grouped based on their contribution ranking (based on the number of times they were selected by the scheduling layer), and these groups are then incrementally scheduled as needed to confidently complete downstream tasks. In this example, the highest ranked group including device A is scheduled first, and the information received from device A is processed by the decoding DNN, as shown in Fig.17 If the information received from device A enables the decoder to complete the downstream task with sufficient confidence, the scheduler does not need to schedule any other groups. On the other hand, if the information from device A is insufficient, the scheduler can then schedule the second highest ranking group including device B and device C. In this scenario, the information previously received from device A can be fused and processed by the decoding DNN with the newly received information from devices B and C, as shown in FIG. Fig.17 If the information received from devices A, B, and C enables the decoder to complete the downstream task with sufficient confidence, the scheduler does not need to schedule the third highest ranked group including device D. Otherwise, the scheduler can then schedule the third highest ranked group. In this case, the information previously received from devices A, B, and C can be fused and processed by the decoding DNN with the newly received information from device D, as shown in FIG. Fig.17 Refer to the instructions at 1206.

[0216] The network prepares the training dataset over a relatively long period of time. The different distributions of adverse channels embedded in the training dataset are usually large-scale due to tall buildings, mountains, trees, etc. Some medium-scale randomness may not be fully represented. For example, a moving truck may block some line of sight (LoS) wireless paths. To handle medium-scale channel randomness, the scheduler can utilize relay nodes. For example, when a contributing device's channel is adversely affected by a medium-scale adverse channel, the contributing device can pass its information to a less contributing device that is less affected by the adverse channel, so that the less contributing device relays the information to the network.

[0217] Fig.18 An example of a less contributing device acting as a relay device for a more contributing device that experiences poor channel conditions is shown. In this example, devices A and C have been identified as contributing to downstream tasks related to object 904. On the other hand, device E is not in a favorable position to observe object 904 and has therefore been determined to be non-contributory to the particular downstream task. At the depicted moment, devices A and E are experiencing favorable channel conditions that enable the devices to send information to network device 452 with an acceptable packet loss rate, as indicated at 1300 and 1306, respectively. However, device C is currently experiencing unfavorable channel conditions, such that the device may be able to receive downlink transmissions from network device 452, but may not be able to successfully send uplink transmissions to network device 452, as indicated at 1302. In this scenario, network device 452 can instruct contributing device C to relay its information to less contributing device E, as indicated at 1304, and allocate uplink wireless resources to device E to send the relayed information from device C to network device 452.

[0218] Training and testing dataset preparation

[0219] In order to prepare a robust training and test data set, the network scheduler may allocate wireless resources among all candidate devices for a period of time so as to sample as many raw measurement results as possible from all candidate devices during the period of time. However, this is usually impractical for several reasons. First, in practice, it is usually too costly to allocate wireless resources to all devices at the same time for a long time. Alternatively, the network scheduler can alternatively randomly sample a certain percentage of N devices, such as 70%, in each sampling interval. In multiple sampling intervals, random sampling can cover all N devices. Secondly, some devices that suffer from severe adverse channel conditions may not be able to successfully send their measurement results to the network even if they are allocated wireless resources. The information or packets of these devices are erased or lost with a certain probability, which can be referred to as the erasure rate or packet loss rate. Third, some devices may take too long to complete their physical measurements / observations. For example, even if wireless resources are allocated within a given sampling interval, the measurement result data of the device may not be ready for transmission. Therefore, each sample of the training and test data sets may not be a full-dimensional (N) sample. The adverse channel conditions (path loss) and measurement result readiness status of the device are embedded in the training and test data sets. For example, although a device occupying a favorable position to measure an object may have a large amount of information, the channel condition of the device may be very poor (e.g., in the shadow of a tall building), so the information (packets) of the device will not appear frequently in the training and test data sets. Therefore, the training may judge that the contribution of the device is small.

[0220] According to the downstream tasks, the sampled raw data can be labeled or grouped to form the final training and testing datasets. Now, the network can start training the DNN so that the scheduler picks a subset of the K devices with the largest contributions and the decoder fuses the received information of the K devices scheduled for the downstream tasks.

[0221] train

[0222] Reference again Fig.13 , it is important to note that the DNN or autoencoder 1100 consists of a scheduling layer or selector layer and several decoding layers. The scheduling layer is the first layer (linear and fully connected from N inputs to K outputs), while the decoding layers constitute the remaining layers. The decoding layers may include one or more fully connected layers, one or more convolutional layers, one or more recursive layers with non-linear activation functions, a combination of one or more of the above types of layers, and so on. As previously discussed, a hyperparameter called temperature adjusts the polarization of the coefficients on the scheduling layer. More details on polarization are discussed in the next section. On this N to K linear fully connected layer, each output is a linear weighted combination of all N inputs.

[0223] The weights on the scheduling layer can be updated by SGD via the back-propagation algorithm, which means that the training objective (downstream task) is considered through the polarization of the scheduling layer.

[0224] Polarization is achieved by the expected training result that for any output, only one of the N weights is very close to 1, and the other N–1 weights are close to 0. For example, w i,j is from the input x i To output y j The weight coefficient of i,j is very close to 1, then input x i is selected for the output y j ; otherwise (w i,j is very close to 0), then input x i Not selected for the output y j .w i,j =0 and w i,k = 1 is possible, which means that the input x i Not selected for the output y j , but is selected for output y k .w i,j =1 and w i,k = 1 is also possible, which means that the input x i Also selected for output y j and output y k In the latter scenario, device i is selected twice.

[0225] As mentioned before, the number of times an input is selected indicates its contribution to the task. The network can score and group the K selected devices according to the number of times each device is selected in the scheduling layer.

[0226] The network generally does not know which value of K is appropriate for a particular downstream task. As mentioned earlier, different tasks may have different appropriate K. The network can use multiple DNNs in parallel with different K, but with the same training and test datasets and the same training objectives. By observing the performance of surviving (scheduled) devices selected more than once by the scheduling layer, the network can determine the minimum number of devices K to schedule. min To achieve training goals.

[0227] Alternatively, the network can start with a larger value of K. In this case, there may be many surviving (scheduled) devices that the scheduling layer has selected more than once. The scheduler can then schedule those devices that have been selected more than once and can perform group-based incremental scheduling as described previously.

[0228] Scheduling

[0229] The weights close to 1 on the scheduling layer indicate to the scheduler which K devices to schedule. The information received from the K devices can then be fused into the decoding layer to complete the task as DNN inference. The scheduling process corresponds to the inference of the trained DNN. Based on the weights close to 1 on the scheduling layer, the network scheduler will allocate uplink radio resources to these K devices (or to fewer devices if the scheduling layer selects some devices more than once) by sending control messages or signaling to these devices. The information received from the scheduled devices will be input to the decoder for processing by the trained decoding layer.

[0230] In some cases, the K devices can be scored or ranked and grouped according to the number of times each device is selected at the scheduling layer. Devices or groups or clusters with higher scores or rankings may contribute more than devices or groups or clusters with lower scores or rankings. Based on the scores or rankings, the scheduler can allocate wireless resources proportionally. In some aspects, this can be regarded as a proportional-fairness (PF) of the contribution of the device rather than the request of the device. For example, the total wireless resources can be allocated among different groups in proportion according to the scores (contributions) of different groups.

[0231] The information received from all scheduled devices will form a K-dimensional input vector for the decoding layer. If some devices do not successfully send their information to the network due to unfavorable channel conditions or unprepared measurement results, the scheduler can replace the information previously received from these unsuccessful devices with either only noise or fixed values. Devices selected more than once can copy their information to several inputs of the decoder. Since groups with higher scores can be given higher priority for radio resource allocation, their information is more likely to successfully reach the network.

[0232] As mentioned earlier, the network scheduler can choose to schedule the highest scoring group as the primary batch, group, or cluster. If the information received from the first batch provides enough confidence for the decoding layer to complete the downstream task, the remaining devices do not need to be scheduled; otherwise, the scheduler can continue to schedule the second highest scoring group so that the information received from the secondary group can be fused with the information from the primary group into the decoding layer to enhance the confidence of the decoder result. This is an opportunistic way to exploit channel diversity and contribution diversity.

[0233] The channel may be subject to medium-sized changes. For example, a moving truck will cause some medium-sized changes in channel conditions. Therefore, a device that is very contributing or has a lot of information during the training phase may suddenly suffer from deteriorating channel conditions during the scheduling (inference) phase. As previously described, in some embodiments, the scheduler can instruct or request this device to pass its measurement results to other devices (e.g., devices with less information but good channel conditions), and instruct or request other devices to act as relay nodes to relay information from the first device to the network.

[0234] Polarization at the Scheduling Layer

[0235] The scheduling layer is a selector based on concrete random variables, which are continuous distributions called concrete distributions. The samples of concrete random variables are:

[0236]

[0237] in,

[0238] w i,j is input from N×1 (x i ) to K×1 output (y j )’s N×K coefficient matrix;

[0239] g i,j is a sample from the Gumbel distribution;

[0240] T is the temperature hyperparameter.

[0241] When T→0, output j, its specific distribution (w i,j The jth column of ) will smoothly approach an N×1 discrete distribution with probability Output i,j = 1 one-hot vector.

[0242] The introduction of concrete distribution brings three significant advantages:

[0243] - Discrete distributions are not differentiable, so SGD backpropagation cannot propagate to the scheduling layer, while the specific layers are mathematically differentiable.

[0244] - The scheduling layer produces a selector. During inference (scheduling), the scheduling layer simply selects K inputs from N inputs based on the weights or coefficients close to 1 on the trained scheduling layer. Therefore, for output y j , instead of observing w i,j Which element of the jth column of is closest to 1, we can select the input α with the largest value i,j ,i=1,2,…,N: This is called the reparameterization trick. In fact, backpropagation will adjust α i,jInstead of directly adjusting w i,j According to α i,j , g i,j and T (current temperature) to calculate w i,j .

[0245] The degree of locality is adjusted by the temperature T. At the beginning of training, T can be initialized to be high enough so that all inputs have an equal chance of being selected for an output. As training progresses, T can be decreased so that only one of the N inputs is selected for this output. For example, T init =10 and T end = 0.01. It should be noted that the specific random distribution does not allow T = 0 but is infinitely close to 0.

[0246] During the training process, T decreasing is called cooling. There are many alternative ways to reduce cooling: linear, exponential decay, etc. Since T decays exponentially in most historical annealing optimizations, it has been used in the following experiments, but aspects of the present invention are by no means limited to this specific cooling technique. The temperature (T) of the current epoch is

[0247]

[0248] Among them, T init is the initial temperature, T end is the end temperature, Epoch index is the index of the current period, N epoch is the total number of epochs. Of course, implementations could have different reduction methods, or even ones based on lookup tables.

[0249] experiment

[0250] In the experiments discussed below, the Modified National Institute of Standards and Technology (MNIST) handwritten digit dataset has been used as an example to demonstrate the effectiveness of the proposed scheduling techniques disclosed in this paper. Although MNIST is a 28×28 image, for the purpose of these experiments, it is treated as a square area with 784 IoT sensors uniformly deployed in a 28×28 array over it. For example, Fig.19 An example is shown of 784 sensors or devices evenly deployed in a 28x28 array to observe area 1404.

[0251] Each sensor measures a small portion of the area (such as a pixel) as 8 bits. Ideally, a total of 784 sensors will send their measurements to the network without error at each time interval to fuse the 784-dimensional input measurements. The application decoder (fuser or processor) at the network device 452 or other location on the network side fuses and processes these 784-dimensional input measurements for three different tasks: classifying the current state of the area corresponding to the 10 numbers in MNIST, reconstructing the reconstructed global picture corresponding to the 28×28 MNIST image, and detecting abnormal states corresponding to the even-odd-number separation (as if even numbers are normal and odd numbers are abnormal).

[0252] Unlike traditional autoencoders for classification and reconstruction, random erasure rates were added independently on each sensor (equivalent to pixels in MNIST) in a uniform pattern to simulate packet loss on the wireless channel 1402 between the device / sensor and the network device 452. Both training and inference (scheduling) will be performed at these random erasure rates, which mimic what would happen when the sensors are connected by a path loss wireless channel. In some later experiments discussed below, some random erasure rates were added in some non-uniform patterns to simulate channel shadowing.

[0253] Fig. 20 The setup of MNIST image input, the introduction of erasure channel probability, and the configuration of a specific encoder of size K and a decoder for the following experimental purposes are shown. Specifically, Fig. 20 As shown, the effect of the random erasure channel is introduced into the original MNIST image pixel data, and the resulting 784 pixel image data input (N=784) is input to a specific encoder 1500 of size K, which varies as discussed in further detail below. The K outputs from the decoder 1500 are then fused and processed in the decoder 1502. The decoder 1502 may include various types of decoding layers, such as one or more fully connected layers, one or more convolutional layers, one or more recursive layers with non-linear activation functions, a combination of one or more of the above types of layers, and the like.

[0254] Fig.21 The specific implementation of the decoder 1502 for the following experimental purposes is shown. Specifically, Fig.21 The decoder 1502 shown includes a first convolutional layer 1602, a second convolutional layer 1604, and a fully connected layer with a nonlinear RelU activation function of size 512. The first convolutional layer 1602 includes 8 one-dimensional filters and the second convolutional layer includes 16 one-dimensional filters. It should be understood that this is only a non-limiting example of a decoder structure.

[0255] Fig. 22Shown is the MNIST handwritten digit input image used in the following experiments.

[0256] Example: With uniform erasure probability E p Categories

[0257] In this experiment, we will study and compare different numbers K of the most contributing devices and various erase rates E in a uniform pattern. p The performance of , in which uniform mode, each sensor (pixel) is independently subjected to the same erasure rate. The task is to find the 10-MNIST digit classification by keeping together the K most contributing devices with various erasure channels from 0% to 90%.

[0258] Fig.23 A plot of simulated accuracy percentage versus erasure probability percentage for a 10-class classification task for different values ​​of encoder size output K is shown. From the simulation results, retaining 128 devices (sensors) (about 1 / 6) out of a total of 784 devices is good enough with an erasure probability of up to 30%. This means that if each sensor has a 30% random packet loss rate on the wireless channel, the scheduler can retain 1 / 6 of the sensors instead of all devices to complete the classification without significantly degrading the performance. Furthermore, if each sensor is affected by a 60% random packet loss rate, the scheduler will have to retain 256 devices, about 1 / 3 of all devices, to compensate for the channel degradation. If the channel continues to degrade, the performance of the classification will suffer regardless of K.

[0259] Example: With uniform erasure probability E p Reconstruction

[0260] In this experiment, we will study and compare different numbers K of the most contributing devices and various erasure channels E. p The task is to reconstruct a 28×28 MNIST image by keeping together the K most contributing devices with various erased channels ranging from 0% to 90%. To compare the performance, the reconstructed image is input to a pre-trained classifier. The performance of the classification will indicate the performance of the reconstruction.

[0261] Fig.24 The reconstructed MNIST images with K = 128 / 256 for different erasure probability percentages are shown. p =30%) and (K=256 and E p =30%) are very similar. We have to input them into the same classifier and measure their classification error in order to compare the two.

[0262] Fig.25A plot of simulated accuracy percentage versus erasure probability percentage for the reconstruction task for different values ​​of the encoder size output K is shown. From the simulation results, retaining 128 devices (sensors) out of a total of 784 devices (about 1 / 6) is good enough with an erasure probability of up to 30%. This means that at a uniform erasure probability of 30% on the wireless channel, the scheduler can retain 1 / 6 of the sensors instead of all the devices to complete the reconstruction. Furthermore, if the channel deteriorates to 60%, the scheduler will have to retain 256 devices, about 1 / 3 of the devices, to compensate for the channel degradation. If the channel continues to degrade, the performance of the reconstruction will suffer regardless of K.

[0263] Example: With uniform erasure probability E p Detection

[0264] In this experiment, we will study and compare different numbers K of the most contributing devices and various erasure channels E. p The task is to detect even-odd numbers by keeping the K most contributing devices with various erasure channels ranging from 0% to 90%.

[0265] Fig.26 A graph showing the simulated accuracy percentage versus erasure probability percentage for the detection task for different values ​​of the encoder size output K. From the simulation results, it seems that retaining 64 devices (sensors) (about 1 / 12) out of a total of 784 devices is good enough with an erasure probability of up to 30%. This means that at a uniform erasure probability of 30% on the wireless channel, the scheduler can retain 1 / 12 of the sensors instead of all devices to complete the detection. Furthermore, if the channel deteriorates to 60%, the scheduler will have to retain 128 devices, about 1 / 6 of the devices, to compensate for the channel degradation. If the channel continues to degrade to 80%, the scheduler will retain 256 devices (1 / 3) to prevent channel degradation.

[0266] More importantly, experiments show that sparsity is task-dependent. The scheduler retains much fewer devices for detection than for classification / reconstruction.

[0267] Example: With E p =0 for the minimum K of the classification task

[0268] Assuming there are no erasures (packet losses) in the wireless transmission, we can study the minimum number of contributing devices (required sparsity) to retain for the classification task. As we mentioned earlier, some devices (or pixels) will be selected multiple times. Then, K min is defined as the number of contributing devices that will be retained to avoid multiple selections.

[0269] Fig. 27Figure 2 shows different values ​​of encoder size output K and 0% erasure probability percentage for a classification task, for a selected set of devices in a 28x28 array. Fig. 27 As shown, when K is 128 and E p = 0, repeated selection occurs (7 to 9 devices are selected more than once). This means that under perfect channel conditions, K for this classification task min This also means that if more than 128 devices are selected (where E p =0), no further classification gain will be obtained. In the ideal transmission case (E p =0), the sparsity associated with this classification task is 1 / 6.

[0270] Of course, if the transfer is not ideal at some non-zero erasure rate, then K min will be increased to compensate for packets lost over the wireless channel. For example, in the diagram showing different values ​​of encoder size output K for a classification task and a 50% erasure probability percentage, in a selected set of 28 devices in a 28x28 array, when E p =50%, K min Increased from 128 to 256 to compensate for packet loss.

[0271] Example: With E p = 0 for the minimum K of the reconstruction task

[0272] Assuming that there are no erasures on the wireless transmission, we can study the minimum number of contributing devices (required sparsity) to be retained for the reconstruction task. As we mentioned earlier, some devices (or pixels) will be selected multiple times. Then, K min is defined as the number of contributing devices that will be retained to avoid multiple selections.

[0273] Fig.29 Figure 2 shows different values ​​of encoder size output K and 0% erasure probability percentage for the reconstruction task, for a selected set of devices in a 28x28 array. Fig.29 As shown, when K is 256 and E p = 0 when repeated selection occurs ( 1 devices are selected more than once). This means that K min will approach 256 This also means that if you select more than 256 Devices (including E p = 0 ), no further reconstruction gain will be obtained. In the ideal transmission case ( E p = 0 ), the sparsity associated with this reconstruction task is 1 / 3. And the K of the classification task min = 128, the results again prove that sparsity is task-dependent. Note that since their compressed sensing, lasso, and PCA tasks are always reconstructions, the sparsity they exploit may not be minimal for tasks simpler than reconstruction.

[0274] Example: With E p =0 for the minimum K detection task

[0275] Assuming that there are no erasures on the wireless transmission, we can study the minimum number of contributing devices (required sparsity) to be retained for the detection task (even-odd separation). As we mentioned earlier, some devices (or pixels) will be selected multiple times. Then, K min is defined as the number of contributing devices that will be retained to avoid multiple selections.

[0276] Fig.30 Figure 2 shows different values ​​of encoder size output K and 0% erasure probability percentage for the detection task, for a selected set of devices in a 28x28 array. Fig.30 As shown, when K is 64 and E p = 0, repeated selection occurs (a device is selected more than once). This means that the K min This also means that if more than 64 devices are selected (where E p =0), no further detection gain will be obtained. In the ideal transmission situation (E p =0), the sparsity associated with this reconstruction task is 1 / 12. min = 128 and K of the reconstruction task min = 256, the results again prove that sparsity is task-dependent.

[0277] Example: In a non-uniform E p The classification task accuracy under the condition of

[0278] Non-uniform E is achieved in situations where obstacles such as adverse terrain affect the erasure probability of a subset of the sensor's transmission channels. p .exist Fig.31 The channel used in the experiment is depicted in , and the effect of this channel on the input can be seen in Fig.32 This gives an overall average

[0279] This non-uniform E p The accuracy of the classification task on the channel and the uniform E p The accuracy of the channels with 0%, 10% and 30% is compared. Fig.33 Shown in.

[0280] The results show that in the case of non-uniform E p The accuracy on the channel (average 30%) is the same as the average 10% E p The accuracy across channels is very similar and much better than the average 30% E p channel, which proves that the scheduler can learn to avoid high E p obstacle area while still maintaining accuracy. This is achieved by studying Fig.34 The pixel selection of the scheduler in the middle is further demonstrated, where it can be seen that the pixels from high E are avoided before the high redundancy scene. p The pixels in the region. It should be noted that the observed K min Close to 128, which is consistent with the uniform E p = 0% of the channel observed K min Consistent.

[0281] Example: In a non-uniform E p The reconstruction task accuracy under the condition of

[0282] This non-uniform E p The accuracy of the reconstruction task on the channel is similar to the uniform E p The accuracy of the channel is compared with that of the channel with 0% and 10%. Fig.35 In addition, Fig.36 It is shown that for the reconstruction task Fig.31 The encoder size output K on the non-uniform erasure channel is shown for different values ​​of,for a selected set of devices in a 28x28 array. Fig.37 Shown in Fig.31 Reconstructed MNIST images with K=32 / 128 / 256 / 512 on the non-uniformly erased channels shown.

[0283] As in the classification task, the results show that p The reconstruction task accuracy on the channel is comparable to the average 10% E p The reconstruction task accuracy on the channel is very similar and much better than the average 30% E p channel, which proves that the scheduler can learn to avoid high E p obstacle area, while still maintaining accuracy. This is again achieved by studying Fig.36 The pixel selection of the scheduler in the middle is further demonstrated, where it can be seen that the pixels from high E are avoided before the high redundancy scene. p The pixels in the region. It should be noted that the observed K min Close to 256, which is consistent with the uniform E p = 0% of the channel observed K min Consistent.

[0284] Example: In a non-uniform E pThe detection task accuracy under the condition of

[0285] This non-uniform E p The accuracy of the detection task on the channel and the uniform E p The accuracy of the channel is compared with that of the channel with 0% and 10%. Fig.38 In addition, Fig.39 It shows that for the detection task Fig.31 The encoder size output K on the non-uniform erasure channel is shown for different values ​​of,for a selected set of devices in a 28x28 array.

[0286] The results show that in the case of non-uniform E p The detection accuracy on the channel is similar to the average 10% E p The detection accuracy on the channel is very similar and is better than 30% E p channels, except when K < 32, in which case the accuracy is slightly affected. The results show that the scheduler is able to learn to avoid high E p obstacle areas while still maintaining accuracy.

[0287] This is again through research Fig.39 The pixel selection of the scheduler in the middle is further demonstrated, where it can be seen that the pixels from high E are avoided before the high redundancy scene. p The pixels in the region. It should be noted that the observed K min Close to 64, which is consistent with the uniform E p = 0% of the channel observed K min Consistent.

[0288] Example: Classification tasks, designing a scheduler to support repeated incremental transfers

[0289] The goal is to implement a HARQ-like scheme where the scheduler first sends a small number of critical samples and then sends more samples if some decision metric fails to meet the threshold.

[0290] ranges from 4 to 512, and in each case the scheduler is trained with constraints to ensure:

[0291] 4≤a <b≤512

[0292] For each K n , where n∈{4,…,512}, train a new decoder on the desired task. p The accuracy of the classification task using this “HARQ training” is compared with the accuracy achieved so far using the “independent training” approach on a channel with 0% Fig.40 In addition, Fig.41Shown are different values ​​of encoder size output K and 0% erasure probability percentage for a classification task under HARQ training for a selected set of devices in a 28x28 array.

[0293] The results show that the accuracy achieved with HARQ training and standalone training is very similar, demonstrating that the scheduler is able to learn the critical set in an incremental manner while still maintaining accuracy. Fig.41 The pixel selection of the scheduler proves that the critical set satisfies the HARQ condition:

[0294] 4≤a <b≤512

[0295] In uniform E p The accuracy of the reconstruction task using this “HARQ training” is compared with the accuracy achieved so far using the “independent training” approach on a channel with 0% Fig.42 Shown in. Fig.43 Figure 2 shows different values ​​of encoder size output K and 0% erasure probability percentage for the reconstruction task under HARQ training, for a selected set of devices in a 28x28 array.

[0296] This demonstrates that the scheduler is able to learn the critical set in an incremental manner while still maintaining accuracy. Fig.43 The scheduler's pixel selection proves that the critical set satisfies the HARQ condition.

[0297] Example: Detection tasks, designing a scheduler to support repeated incremental transfers

[0298] In uniform E p The accuracy of the detection task using this "HARQ training" is compared with the accuracy achieved so far using the "independent training" method on a channel with 0% detection. Fig.44 Shown in. Fig.45 Shown are different values ​​of encoder size output K and 0% erasure probability percentage for the detection task under HARQ training for a selected set of devices in a 28x28 array.

[0299] The results show that the accuracy achieved with HARQ training and standalone training is very similar, demonstrating that the scheduler is able to learn the critical set in an incremental manner while still maintaining accuracy. Fig.45 The scheduler's pixel selection proves that the critical set satisfies the HARQ condition.

[0300] In wireless systems, many devices (terminals, device-devices, etc.) are connected to the network through one or several base stations (BTS, eNode, access points, etc.). These devices are measuring, observing, and collecting information about natural phenomena (objectives, targets, etc.). The network has a scheduler to allocate UL radio resources for these devices to send their measurements or observations back to the network.

[0301] Uplink (UL) radio resources can be allocated in terms of bandwidth, modulation and coding scheme, packet size, transmission duration, and / or spreading codes. Basically, the more radio resources a device gets, the more likely it is to successfully send its measurements to the network. Scheduling messages are sent on downlink channels (control channels or dedicated data channels). Since the BTS can have much higher transmission power, we assume that the scheduling messages can successfully reach the scheduled devices in a timely manner.

[0302] However, due to the limitation of UL radio resources and different (non-uniform) path losses among devices, not every device can obtain UL radio resources when requested, and not all devices can send their measurements to the network in time. Therefore, the scheduler must carefully select the device and allocate UL radio resources to the device in this transmission interval. Traditionally, the scheduler will use the PF algorithm, which will ignore the different path losses among devices and allocate radio resources to the device in proportion to the request of this device.

[0303] Very different from the traditional PF scheduler, the contribution-based scheduler disclosed in this article will schedule devices (select devices and allocate UL radio resources to devices) according to their contribution to downstream tasks at the network side. The contribution index not only indicates how much information the device has for the task, but also indicates to what extent the device sends this measurement result to the network.

[0304] As mentioned before, a device has little idea of ​​how much it contributes to the global task on the network. Therefore, the device will assume that it contributes as much as other devices. This assumption will lead to such a homogeneous request distribution among devices observing a common goal. This high entropy multi-device request distribution brings little system gain to the request-based PF scheduler.

[0305] Contribution-based scheduling will eliminate the devices with the smallest contribution, either due to their unfavorable observation locations, or due to their severe path loss, or both. It would be a waste to allocate radio resources to these devices with the smallest contribution.

[0306] In addition, in the contribution-based scheduling method disclosed herein, the contribution is specific to a given downstream task. The contribution of a device may vary from task to task. A device that contributes a lot to one task may find itself irrelevant to another task.

[0307] In fact, multiple tasks can be completed in parallel by a group of IoT devices. Each task will identify its associated contributing devices. Some devices may contribute to more than one task; some devices may contribute to only one task; and some devices may not contribute to any task.

[0308] The contribution of the device is learned through the original dataset (training and test datasets) and the specific task. The learning is performed by an autoencoder with at least two layers. The first layer is a linear fully connected layer that acts as a scheduling layer or a selector layer; the remaining layers can be linear, non-linear, convolutional layers, etc. that act as decoding layers. The training objectives of this autoencoder are related to the downstream tasks: classification, detection, reconstruction, long-term reward expectation for reinforcement learning, etc.

[0309] Before the training phase, the network will prepare training and test datasets. In most cases, the network will evenly and randomly allocate resources for a certain percentage of devices to send their raw measurements back to the network at a time interval. Due to different packet loss rates of different devices to BTS channels, some packets may not reach the network successfully or in a timely manner. These lost packets will be recorded in the training and test datasets to reflect the current path loss distribution among the devices. The network will collect raw measurements over a sufficient number of time intervals.

[0310] After collecting a sufficient amount of training and testing data, the network can train the autoencoder to learn one or more of the following:

[0311] - Equipment that contributes most to the mission;

[0312] - The minimum number of devices K to schedule for the task min ;

[0313] - Clusters of contributing devices in terms of contribution.

[0314] As mentioned before, training can be based on SGD backpropagation from the last layer to the first layer (scheduling layer or selector layer) so that each neuron will be exposed to the training target (task).

[0315] After training is complete, the network's scheduler can schedule devices based on the trained weights or coefficients of the scheduling layer. The scheduled devices send their measurements to the network, which inputs the measurements into the decoding layer.

[0316] This is a nondeterministic polynomial (NP) problem for jointly optimizing the information content index of the device used for the task and the conditions of the channel connection between the device and the network. There are various tasks and various channel conditions. The advantage of DNN is that back propagation (SGD) can propagate the task (training target) from the last layer to the first layer. Then, all the neurons of this DNN work together to complete the task. The first layer will adjust the scheduler, and the remaining layers will fuse and process the incoming information from multiple devices. In addition, information about the path loss rate among the devices is embedded in the training and test data sets. Because it is data-driven, this channel path loss factor will be implicitly taken into account in the optimization of DNN. The scheduling method based on the autoencoder (DNN) disclosed in this article provides a global optimization platform for joint optimization.

[0317] The scheduling layer is a fully connected linear layer: each of its inputs is linked to each of its outputs. Each output is a weighted linear combination of all inputs. The measurement result information (raw information) from one device is considered as one input (or one input dimension). If there are N devices in total, there are N inputs. The scheduler will select the K devices with the largest contribution from the total N candidates, which corresponds to the K inputs retained and processed by the subsequent decoding layers of the decoder part of the autoencoder.

[0318] Unlike a traditional linear layer, training polarizes this layer. At the end of training, although the output is a weighted linear combination from N inputs, only one of the N weights is close to 1, and the rest are close to 0. This indicates that inputs with weights close to 1 are selected for this output. At the end of training, the scheduling layer becomes an N to K selector.

[0319] To polarize the first layer for training, the discrete distribution is replaced by a concrete distribution. The concrete distribution is parameterized by a temperature. As training progresses (epoch by epoch), the temperature gets colder, making the weights of the linear combination more and more polarized.

[0320] The benefit of using a concrete distribution instead of a discrete distribution is that the latter is differentiable for SGD. As the temperature approaches zero, the concrete distribution will be very similar to the discrete distribution. The concrete distribution at low temperatures will ensure that only one of the N weights is close to 1, and the rest are close to 0.

[0321] The minimum number of devices to be scheduled K min Depends on the task. Different tasks may have different K min In order to obtain the task K min , we can try different K under the same training and test datasets and the same training objective.

[0322] K minis the minimum K that not only meets the training objective but also does not cause any device to be selected more than once.

[0323] On a schedule layer trained with a specific distribution, an input is allowed to be selected by more than one output. This means that although the AE specifies K outputs, the trained schedule layer may indicate less than K input devices to be selected because some selected devices are selected for more than one output.

[0324] K min Devices make up the contributing set for a task. This means that the devices in the contributing set will be sufficient to complete the task (including the standard). For example, in a classification task, a larger contributing set is required to achieve higher classification accuracy and vice versa.

[0325] Task K min Divide by N(R min =K min / N) is similar to the determination of compression rate, sampling rate and channel rate. From the perspective of information theory, compression rate, sampling rate and channel rate are to find the typical set related to the reconstruction task using the error criterion (square error or bit error). The typical set assumes that the channel is stationary, either as an independent stationary AWGN and binary erasure channel or as a stationary Markov chain. However, our R min =K min / N is to find the contribution set associated with any task using any criteria in different non-uniformly erased channels. Therefore, the contribution set seems to be more general than the typical set.

[0326] Due to K min Associating tasks (and their criteria) with channel conditions, the scheduler can cluster or group selected devices based on the number of times they were selected. min ,Some devices may be selected more than once. The number of the selected device indicates its contribution to the task (and channel).

[0327] Both the observed phenomena and the channels vary randomly, which leads to some differences in the measurements and channel conditions. To benefit from the diversity of opportunities, the scheduler can first schedule the devices of the primary group that are selected the most times. If the information from the primary group successfully reaches the network and provides enough confidence for the decoder to complete the task, the scheduler can avoid scheduling the secondary group. Otherwise, the scheduler can schedule the devices of the secondary group, and the information from the primary and secondary groups will be input to the decoder together to improve its confidence. In some DNN cases, softmax is used to indicate a kind of confidence. For example, for a 10-class classification, if no softmax value reaches a certain threshold, the confidence is low. However, if one of the 10 softmax values ​​is prominent, this can provide high confidence in the classification.

[0328] In practice, there can be one DNN trained using both the primary set of inputs and the primary / secondary set of inputs. Alternatively, there can be one DNN trained using the primary set of inputs and another DNN trained using the primary / secondary set.

[0329] This concept is similar to traditional HARQ-based channel coding schemes. The hybrid incremental redundancy scheduler (based on contribution) will exploit the diversity gain (measurement results and channels). In contrast, HARQ channel coding only exploits the diversity gain of the channel. Therefore, the method disclosed in this article considers more diversity than traditional HARQ schemes.

[0330] Since the channel is varying, some devices in the contributing set may suffer from sudden deep shadowing, although their measurements may be very informative for the task. In this case, an alternative is to schedule relays, where these informative devices will send their measurements through less informative devices that have good channel conditions.

[0331] AI-based sampling algorithm

[0332] Although many of the examples described above relate to AI-based schedulers and scheduling algorithms, aspects of the present disclosure are also applicable to AI-based sampling algorithms, some examples of which are described below.

[0333] The general goal of sampling is to find the most representative samples from which the original information being sampled can be reconstructed. Two commonly used traditional sampling techniques are called Nyquist sampling and compressed sensing. In Nyquist sampling, there is an L2 optimization and there is no distinction between samples. In compressed sensing, there is an L1 optimization and there is no distinction between samples.

[0334] Aspects of the present disclosure provide a smart sampling algorithm that has several advantages over Nyquist and compressed sensing sampling algorithms, such as:

[0335] Find the samples that contribute most to downstream tasks;

[0336] Neither l1 nor l2 are optimized

[0337] There are differences in contribution among the samples

[0338] Information fusion is needed

[0339] Can find suitable sampling sparsity (compressed detection cannot do this, the Nyquist limit is 2xf max )

[0340] Sparsity depends on the downstream task. (Compressed sensing and Nyquist only deal with sparsity and reconstruction).

[0341] Other smart sampling applications

[0342] Interference generation:

[0343] In full duplex, interference is generated by full resolution sampling of the transmitted signal. Aspects of the present disclosure may be utilized to reduce sampling to generate interference, and aspects of the present disclosure may potentially be extended to other interference generating applications.

[0344] Sparse Channel Lease

[0345] Instead of measuring high-resolution wireless channels across an entire cell, aspects of the present invention can be applied to find the most contributing locations to lease cellular channels.

[0346] Bottleneck nodes in self-organizing networks

[0347] Aspects of the present invention may be applied to identify bottleneck network nodes to maintain connections associated with tasks.

[0348] Intelligent Quantification

[0349] Aspects of the invention may be applied to find the minimum quantization for a particular task, ie the minimum number of quantization levels required for a particular task, which may vary significantly from task to task.

[0350] Also disclosed are examples of devices (e.g., ED or UE and TRP or network device) for performing the various methods described herein.

[0351] For example, the first device may include a memory for storing processor-executable instructions, and a processor for executing the processor-executable instructions. When the processor executes the processor-executable instructions, the processor may be caused to perform the method steps of one or more devices described herein. For example, the processor may enable the device to communicate over the air interface in the operating mode by implementing operations consistent with the operating mode, for example, performing necessary measurements and generating content from those measurements (as configured for the operating mode), preparing uplink transmissions and processing downlink transmissions (e.g., encoding, decoding, etc.), and configuring and / or instructing transmission / reception on the RF chain and antenna.

[0352] Note that the expression "at least one of A or B" as used herein is interchangeable with the expression "A and / or B". This expression refers to a list where either A or B or both A and B can be selected. Similarly, "at least one of A, B, or C" as used herein is interchangeable with "A and / or B and / or C" or "A, B, and / or C". This expression refers to a list where either A or B or C can be selected, or both A and B, or both A and C, or both B and C, or all of A, B, and C. The same principles apply to longer lists with the same format.

[0353] Although the present invention has been described with reference to specific features and embodiments of the present invention, various modifications and combinations of the present invention may be formulated without departing from the present invention. The specification and drawings are therefore only considered as an illustration of some embodiments of the present invention as defined by the appended claims, and any and all modifications, variants, combinations or equivalents covered within the scope of the present invention are considered. Therefore, although the present invention and its advantages have been described in detail, various changes, substitutions and modifications may be made without departing from the present invention as defined by the appended claims. In addition, the scope of the present application is not limited to the specific embodiments of the processes, machines, manufactured products, material compositions, components, methods and steps described in the specification. A person of ordinary skill in the art will readily understand from the disclosure of the present invention that a process, machine, manufactured product, material composition, component, method or step currently existing or to be developed that performs substantially the same function as the corresponding embodiment described herein, or that can achieve substantially the same result as the embodiment, may be used according to the present invention. Therefore, the appended claims are intended to include these processes, machines, manufactured products, material compositions, components, methods or steps within their scope.

[0354] In addition, any module, component, or device illustrated herein that executes instructions may include or otherwise access one or more non-transitory computer / processor readable storage media to store information, such as computer / processor readable instructions, data structures, program modules, and / or other data. A non-exhaustive list of examples of non-transitory computer / processor readable storage media includes cassettes, magnetic tapes, disk storage or other magnetic storage devices, such as compact disc read-only memory (CD-ROM), digital video discs or digital versatile discs (DVD), Blu-ray discs, TMCD-ROMs, or other optical storage, volatile and non-volatile, removable and non-removable media implemented in any method or technology, random-access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies. Any of these non-transitory computer / processor storage media can be part of the device or can be accessible or connected to the device. Any application or module described in this article can be implemented using computer / processor readable / executable instructions, which can be stored or otherwise saved by these non-transitory computer / processor readable storage media.

[0355] Abbreviation Definition

[0356] LTE Long Term Evolution

[0357] NR New Radio

[0358] BWP Bandwidth Part

[0359] BS Base Station

[0360] CA Carrier Aggregation

[0361] CC Component Carrier

[0362] CG Cell Group

[0363] CSI Channel State Information

[0364] CSI-RS Channel State Information Reference Signal

[0365] DC Dual Connection

[0366] DCI Downlink Control Information

[0367] DL Downlink

[0368] DL-SCH Downlink Shared Channel

[0369] EN-DC E-UTRA NR dual connectivity, MCG uses E-UTRA, SCG uses NR

[0370] gNB Next Generation (or 5G) Base Station

[0371] HARQ-ACK Hybrid Automatic Repeat Request Acknowledgement

[0372] MCG Master Cell Group

[0373] MCS Modulation Coding Scheme

[0374] MAC-CE Media Access Control-Control Element

[0375] PBCH Physical Broadcast Channel

[0376] Pcell Primary cell

[0377] PDCCH Physical Downlink Control Channel

[0378] PDSCH Physical Downlink Shared Channel

[0379] PRACH Physical Random Access Channel

[0380] PRG Physical Resource Group

[0381] PSCell Main SCG cell

[0382] PSS Primary Synchronization Signal

[0383] PUCCH Physical Uplink Control Channel

[0384] PUSCH Physical Uplink Shared Channel

[0385] RACH Random Access Channel

[0386] RAPID Random Access Preamble Identifier

[0387] RB Resource Block

[0388] RE Resource Element

[0389] RRM Radio Resource Management

[0390] RMSI Remaining System Information

[0391] RS reference signal

[0392] RSRP Reference Signal Received Power

[0393] RRC Radio Resource Control

[0394] SCG Secondary Cell Group

[0395] SFN System Frame Number

[0396] SL Sidelink

[0397] SCell Secondary Cell

[0398] SPS Semi-Persistent Scheduling

[0399] SR Scheduling Request

[0400] SRI SRS Resource Indicator

[0401] SRS Sounding Reference Signal

[0402] SSS Secondary synchronization signal

[0403] SSB Synchronous Signal Block

[0404] SUL Supplementary Uplink

[0405] TA Timing Advance

[0406] TAG Timing Advance Group

[0407] TUE Target UE

[0408] UCI Uplink Control Information

[0409] UE User Equipment

[0410] UL Uplink

[0411] UL-SCH Uplink Shared Channel

Claims

1. A method for scheduling uplink transmission in a wireless communication network, characterized in that: The method comprises: Selecting a first plurality of devices from a set of candidate devices to be scheduled for uplink transmission, the selecting the first plurality of devices being based on a first contribution indicator of each device, wherein the first contribution indicator of each device is associated with a first downstream task in the wireless communication network and indicates: i) the extent to which the device can successfully send information for the first downstream task to the network; ii) how much information the information provided by the device is for the first downstream task; Scheduling information for the first plurality of devices is sent, where the scheduling information for the first plurality of devices indicates uplink radio resources allocated to the first plurality of devices.

2. The method according to claim 1, characterized in that The uplink radio resources are allocated to the first plurality of devices by allocating uplink radio resources to each device in the first plurality of devices based on a first contribution indicator of the device.

3. The method according to claim 2, characterized in that A device having a first contribution indicator indicating a higher contribution to the first task is allocated more uplink radio resources than a device having a first contribution indicator indicating a lower contribution to the first task.

4. The method according to any one of claims 1 to 3, characterized in that Also includes: Selecting a second plurality of devices from the set of candidate devices to be scheduled for uplink transmission, the selecting the second plurality of devices being based on a second contribution indicator of each device, wherein the second contribution indicator of each device is associated with a second downstream task in the wireless communication network that is different from the first downstream task, wherein the second contribution indicator indicates: i) the extent to which the device can successfully send information for the second downstream task to the network; ii) how much information the information provided by the device is for the second downstream task; Scheduling information for the second plurality of devices is sent, where the scheduling information for the second plurality of devices indicates uplink radio resources allocated to the second plurality of devices.

5. The method according to any one of claims 1 to 3, characterized in that The first contribution indicator of the candidate devices for the first downstream task is learned by machine learning using a machine learning module, wherein the machine learning module includes a deep neural network (DNN), and the DNN is trained using raw test data received from at least a subset of the candidate devices as ML module input and one or more parameters for the first downstream task as ML module output to meet a training objective related to the first downstream task.

6. The method according to claim 5, characterized in that The DNN is configured as an autoencoder including at least two layers of neurons, wherein a first layer of the autoencoder is a linear fully connected layer, the linear fully connected layer includes K neurons, the K neurons have N inputs and K outputs corresponding to a set of N candidate devices, each of the K outputs of the first layer is a weighted linear combination of the N inputs, wherein once trained, the first layer of the autoencoder is configured as an N to K selector that selects K inputs from the set of N inputs, wherein K <N。 7. The method according to claim 6, characterized in that One or more layers after the first layer of the autoencoder are configured as a decoder to perform decoding of the first downstream task using the K outputs from the first layer as inputs of the decoder.

8. The method according to claim 6 or 7, characterized in that: The training of the autoencoder is based on stochastic gradient descent (SGD) back-propagation from the last layer of the autoencoder to the first layer of the autoencoder to meet the training objective related to the first downstream task.

9. The method according to any one of claims 6 to 8, characterized in that Selecting the first plurality of devices to be scheduled for uplink transmission based on a first contribution indicator of each device includes: selecting the first plurality of devices based on weights of a trained first layer of the autoencoder.

10. The method according to any one of claims 6 to 9, characterized in that Training of the autoencoder polarizes the weights in a first layer of the autoencoder such that for each of the K neurons in the first layer of the autoencoder, an output of the neuron is a weighted combination of the N inputs to the neuron, but only one of the N weights is close to a value of 1 and the remaining N-1 weights are close to a value of 0.

11. The method according to claim 10, characterized in that The training of the autoencoder utilizes a continuous relaxation of a discrete distribution (specifically a distribution) parameterized by a temperature parameter, wherein the temperature parameter is decreased over the course of multiple training epochs such that the weights of the first layer of the autoencoder become increasingly polarized.

12. The method according to claim 10 or 11, characterized in that: For each of the K neurons in the first layer of the autoencoder, candidate devices corresponding to inputs for which the trained weight of the neuron is close to a value of 1 are considered to have been selected by the neuron.

13. The method according to claim 12, characterized in that The number of neurons K in the first layer of the autoencoder is equal to K of the first downstream task min , where K min is a downstream task-specific value, and K of the first downstream task min A minimum number of neurons in the first layer is identified and indicated during training of the autoencoder for the first downstream task, wherein the neurons enable meeting the training objective associated with the first downstream task without requiring more than one of the neurons in the first layer to select any of the candidate devices.

14. The method according to claim 13, characterized in that determining K for the first downstream task during training of the autoencoder for the first downstream task by training multiple versions of the autoencoder using the same original test data as an input to the autoencoder and the same training objective associated with the first downstream task, but with different numbers of neurons in a first layer of the autoencoder min .

15. The method according to claim 12, characterized in that At least one of the candidate devices is selected by one or more of the K neurons in the first layer.

16. The method according to claim 15, characterized in that The first contribution indicator of each candidate device is based on the number of times the candidate device is selected in the first layer of the autoencoder.

17. The method according to claim 15, characterized in that Also includes: dividing the candidate devices into a plurality of groups based on the number of times the candidate devices are selected in the first layer of the autoencoder, the plurality of groups comprising at least a primary group and a secondary group, wherein candidate devices grouped into the primary group are selected more times in the first layer of the autoencoder than candidate devices grouped into the secondary group, selecting the first plurality of devices to be scheduled for uplink transmission comprises selecting devices of the primary group, Sending uplink scheduling information for the first plurality of devices includes sending primary uplink scheduling information for devices of the primary group.

18. The method according to claim 17, characterized in that Each of the candidate devices grouped into the secondary group is selected at least once in the first layer of the autoencoder.

19. The method according to claim 17 or 18, characterized in that Also includes: receiving an uplink transmission from a device in the primary group of devices according to the primary uplink scheduling information; The uplink transmissions received from the primary group of devices are used as input to a trained decoder to perform decoding of the first downstream task.

20. The method according to claim 19, characterized in that Also includes: determining one or more confidence indicators based on decoding of the first downstream task using the uplink transmission received from the devices of the primary group as input to the trained decoder; A determination is made based on the one or more confidence indicators whether to send secondary uplink scheduling information for devices of the secondary group.

21. The method according to claim 20, characterized in that Determining whether to send secondary uplink scheduling information for the devices of the secondary group based on the one or more confidence indicators includes: after determining that the one or more confidence indicators indicate that the result of decoding the first downstream task using the uplink transmission received from the devices of the primary group as input to the trained decoder has sufficient confidence, determining not to send the secondary uplink scheduling information for the devices of the secondary group.

22. The method according to claim 20, characterized in that The method further comprises, after determining that the one or more confidence indicators indicate that a result of decoding of the first downstream task using the uplink transmission received from the devices of the primary group as input to the trained decoder has insufficient confidence: Sending secondary uplink scheduling information for the devices in the secondary group, wherein the secondary uplink scheduling information indicates, for each device in the secondary group, an uplink radio resource allocated to the device; receiving an uplink transmission from a device in the secondary group of devices according to the secondary uplink scheduling information; The uplink transmissions received from the devices of the primary group and the uplink transmissions received from the devices of the secondary group are used as inputs to the trained decoder to perform decoding of the first downstream task.

23. The method according to any one of claims 20 to 22, characterized in that Determining one or more confidence indicators includes determining one or more softmax values.

24. A network device, characterized in that: include: processor; a memory storing processor-executable instructions that, when executed, cause the processor to: Selecting a first plurality of devices from a set of candidate devices to be scheduled for uplink transmission, wherein selecting the first plurality of devices is based on a first contribution indicator of each device, wherein the first contribution indicator of each device is associated with a first downstream task in a wireless communication network and indicates: i) the extent to which the device can successfully send information for the first downstream task to the network; ii) how much information the information provided by the device is for the first downstream task; Scheduling information for the first plurality of devices is sent, where the scheduling information for the first plurality of devices indicates uplink radio resources allocated to the first plurality of devices.

25. The network device according to claim 24, characterized in that: The uplink radio resources are allocated to the first plurality of devices by allocating uplink radio resources to each device in the first plurality of devices based on a first contribution indicator of the device.

26. The network device according to claim 25, characterized in that: A device having a first contribution indicator indicating a higher contribution to the first task is allocated more uplink radio resources than a device having a first contribution indicator indicating a lower contribution to the first task.

27. The network device according to any one of claims 24 to 26, characterized in that: The processor-executable instructions also include processor-executable instructions that, when executed, cause the processor to: Selecting a second plurality of devices from the set of candidate devices to be scheduled for uplink transmission, the selecting the second plurality of devices being based on a second contribution indicator of each device, wherein the second contribution indicator of each device is associated with a second downstream task in the wireless communication network that is different from the first downstream task, wherein the second contribution indicator indicates: i) the extent to which the device can successfully send information for the second downstream task to the network; ii) how much information the information provided by the device is for the second downstream task; Scheduling information for the second plurality of devices is sent, where the scheduling information for the second plurality of devices indicates uplink radio resources allocated to the second plurality of devices.

28. The network device according to any one of claims 24 to 26, characterized in that: The first contribution indicator of the candidate devices for the first downstream task is learned by machine learning using a machine learning module, wherein the machine learning module includes a deep neural network (DNN), and the DNN is trained using raw test data received from at least a subset of the candidate devices as ML module input and one or more parameters for the first downstream task as ML module output to meet a training objective related to the first downstream task.

29. The network device according to claim 28, characterized in that: The DNN is configured as an autoencoder including at least two layers of neurons, wherein a first layer of the autoencoder is a linear fully connected layer, the linear fully connected layer includes K neurons, the K neurons have N inputs and K outputs corresponding to a set of N candidate devices, each of the K outputs of the first layer is a weighted linear combination of the N inputs, wherein once trained, the first layer of the autoencoder is configured as an N to K selector that selects K inputs from the set of N inputs, wherein K <N。 30. The network device according to claim 29, characterized in that: One or more layers after the first layer of the autoencoder are configured as a decoder to perform decoding of the first downstream task using the K outputs from the first layer as inputs of the decoder.

31. The network device according to claim 29 or 30, characterized in that: The training of the autoencoder is based on stochastic gradient descent (SGD) back-propagation from the last layer of the autoencoder to the first layer of the autoencoder to meet the training objective related to the first downstream task.

32. The network device according to any one of claims 29 to 31, characterized in that: Selecting the first plurality of devices to be scheduled for uplink transmission based on a first contribution indicator of each device includes: selecting the first plurality of devices based on weights of a trained first layer of the autoencoder.

33. The network device according to any one of claims 29 to 32, characterized in that: Training of the autoencoder polarizes the weights in a first layer of the autoencoder such that for each of the K neurons in the first layer of the autoencoder, an output of the neuron is a weighted combination of the N inputs to the neuron, but only one of the N weights is close to a value of 1 and the remaining N-1 weights are close to a value of 0.

34. The network device according to claim 33, characterized in that: The training of the autoencoder utilizes a continuous relaxation of a discrete distribution (specifically a distribution) parameterized by a temperature parameter, wherein the temperature parameter is decreased over the course of multiple training epochs such that the weights of the first layer of the autoencoder become increasingly polarized.

35. The network device according to claim 33 or 34, characterized in that: For each of the K neurons in the first layer of the autoencoder, candidate devices corresponding to inputs for which the trained weight of the neuron is close to a value of 1 are considered to have been selected by the neuron.

36. The network device according to claim 35, characterized in that The number of neurons K in the first layer of the autoencoder is equal to K of the first downstream task min , where K min is a downstream task-specific value, and K of the first downstream task min A minimum number of neurons in the first layer is identified and indicated during training of the autoencoder for the first downstream task, wherein the neurons enable meeting the training objective associated with the first downstream task without requiring more than one of the neurons in the first layer to select any of the candidate devices.

37. The network device according to claim 36, characterized in that: determining K for the first downstream task during training of the autoencoder for the first downstream task by training multiple versions of the autoencoder using the same original test data as an input to the autoencoder and the same training objective associated with the first downstream task, but with different numbers of neurons in a first layer of the autoencoder min .

38. The network device according to claim 35, characterized in that: At least one of the candidate devices is selected by one or more of the K neurons in the first layer.

39. The network device according to claim 38, characterized in that: The first contribution indicator of each candidate device is based on the number of times the candidate device is selected in the first layer of the autoencoder.

40. The network device according to claim 38, characterized in that The processor-executable instructions also include processor-executable instructions that, when executed, cause the processor to: dividing the candidate devices into a plurality of groups based on the number of times the candidate devices are selected in the first layer of the autoencoder, the plurality of groups comprising at least a primary group and a secondary group, wherein candidate devices grouped into the primary group are selected more times in the first layer of the autoencoder than candidate devices grouped into the secondary group, selecting the first plurality of devices to be scheduled for uplink transmission comprises selecting devices of the primary group, Sending uplink scheduling information for the first plurality of devices includes sending primary uplink scheduling information for devices of the primary group.

41. The network device according to claim 40, characterized in that Each of the candidate devices grouped into the secondary group is selected at least once in the first layer of the autoencoder.

42. The network device according to claim 40 or 41, characterized in that: The processor-executable instructions also include processor-executable instructions that, when executed, cause the processor to: receiving an uplink transmission from a device in the primary group of devices according to the primary uplink scheduling information; The uplink transmissions received from the primary group of devices are used as input to a trained decoder to perform decoding of the first downstream task.

43. The network device according to claim 42, characterized in that: The processor-executable instructions also include processor-executable instructions that, when executed, cause the processor to: determining one or more confidence indicators based on decoding of the first downstream task using the uplink transmission received from the devices of the primary group as input to the trained decoder; A determination is made based on the one or more confidence indicators whether to send secondary uplink scheduling information for devices of the secondary group.

44. The network device according to claim 43, characterized in that Determining whether to send secondary uplink scheduling information for the devices of the secondary group based on the one or more confidence indicators includes: after determining that the one or more confidence indicators indicate that the result of decoding the first downstream task using the uplink transmission received from the devices of the primary group as input to the trained decoder has sufficient confidence, determining not to send the secondary uplink scheduling information for the devices of the secondary group.

45. The network device according to claim 43, characterized in that The processor-executable instructions also include processor-executable instructions that, when executed, cause the processor to perform the following operations: upon determining that the one or more confidence indicators indicate that a result of decoding of the first downstream task using the uplink transmission received from the primary group of devices as input to the trained decoder has insufficient confidence: Sending secondary uplink scheduling information for the devices in the secondary group, wherein the secondary uplink scheduling information indicates, for each device in the secondary group, an uplink radio resource allocated to the device; receiving an uplink transmission from a device in the secondary group of devices according to the secondary uplink scheduling information; The uplink transmissions received from the primary group of devices and the uplink transmissions received from the secondary group of devices are used as inputs to the trained decoder to perform decoding of the first downstream task.

46. ​​The network device according to any one of claims 43 to 45, characterized in that: Determining one or more confidence indicators includes determining one or more softmax values.

47. A device, characterized in that Comprising one or more units for performing the method according to any one of claims 1 to 23.

Images