Method, architecture, apparatus and system for online training and selection of neurobelief propagation decoder

Abstract

Processes, methods, architectures, apparatuses, systems, devices, and computer program products are provided for online training and / or selection of a Neurobelief Propagation (NBP) decoder. A method may include receiving training configuration information indicating an index of a Neurobelief Propagation (NBP) variant or architecture and / or an update to one or more received hyper-parameters for the NBP variant or architecture. The method may include transmitting an indication that the WTRU is ready to train an NBP variant or architecture, receiving a training dataset including all-zero codewords associated with a channel coding technique, starting training the NBP variant or architecture using the received training dataset and / or hyper-parameters, and after one or more training iterations, generating a channel coding technique using the NBP variant or architecture. Information indicative of a cost function value associated with the trained NBP variant or architecture is sent. When training is completed, the method may include sending information indicating trained weights associated with the trained NBP variants or architectures.

Description

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS

[0002] This application claims the benefit of U.S. Provisional Patent Application No. 63 / 448,458, filed on February 27, 2023, the entire contents of which are incorporated herein by reference. Technical Field

[0003] Various embodiments described in this disclosure generally relate to the fields of communications, software, and coding, including, for example, methods, architectures, apparatuses, and systems related to online training and / or selection of neural belief propagation (NBP) decoders. Background Art

[0004] Neural Belief Propagation (NBP) decoding is a scheme that can decode different error-correcting code families. NBP decoding can be considered as a hybrid of deep neural network (DNN) and belief propagation (BP) decoding. Summary of the Invention

[0005] Embodiments may be directed to a method implemented in a wireless transmit / receive unit (WTRU). The method may include receiving training configuration information from a network element, the training configuration information indicating (i) an index of a neural belief propagation (NBP) variant or architecture and / or (ii) updates to one or more received hyperparameters for the NBP variant or architecture. The method may include sending first information to the network element indicating that the WTRU is ready to train the NBP variant or architecture; and receiving a training dataset including all-zero codewords associated with a channel coding technique from the network element. The all-zero codewords may have a defined information block length and code rate. The method may then include starting training the NBP variant or architecture using the received training dataset and any of the hyperparameters. After one or more training iterations, the method may include sending second information to the network element indicating a cost function value associated with the trained NBP variant or architecture, and upon completion of the training, sending third information to the network element indicating trained weights associated with the trained NBP variant or architecture.

[0006] One embodiment can involve a WTRU that includes circuitry (e.g., a processor, memory, and / or transceiver) configured to receive, from a network element, training configuration information indicating (i) an index of a neural belief propagation (NBP) variant or architecture and / or (ii) an update to one or more received hyperparameters for the NBP variant or architecture. The WTRU can be configured to transmit, to the network element, first information indicating that the WTRU is ready to train the NBP variant or architecture and receive, from the network element, a training dataset including an all-zero codeword associated with a channel coding technique. The all-zero codeword can have a defined information block length and code rate. The WTRU can be configured to begin training of the NBP variant or architecture using the received training dataset and any of the hyperparameters. After one or more training iterations, the WTRU can be configured to transmit, to the network element, second information indicating a cost function value associated with the trained NBP variant or architecture and, if the training is complete, third information indicating trained weights associated with the trained NBP variant or architecture.

[0007] In various embodiments, the method can include, or the WTRU can be configured to, receive configuration information indicating a lookup table associating NBP models and parameters.

[0008] In various embodiments, the NBP models can include FNSPA, RNSPA, FNNMS, RNNMS, NBP-SS, and / or NBP-SS-PAN.

[0009] In various implementations, the method can include, or the WTRU can be configured to, receive, from the network element, a training indicator, where the training indicator includes any of an indication of a number of OFDM symbols, slots, and subframes before the training begins.

[0010] In various implementations, the method can include, or the WTRU can be configured to, transmit, to the network element, a request to receive training configuration. In some embodiments, the second information can be transmitted after each training iteration, after a predefined set of iterations, and / or when the training is complete.

[0011] In various embodiments, the training is completed (e.g., deemed complete) after all training iterations are completed or on a condition that a loss function of the NBP variant or architecture exceeds a predefined threshold.

[0012] In various embodiments, the method can include or the WTRU can be configured to receive a BLER reference signal (BRS) from a network element, where the BRS can indicate a pair of SNR / BLER values and / or an index of the NBP. In one embodiment, the method can include (or the WTRU can be configured to) compare the SNR / BLER values to those in a stored BLER table, determine whether the weights of the NBP model should be updated based on the comparison, and transmit an indication of the update to the model to the network element if the weights are updated.

[0013] In various embodiments, the training can be performed in cooperation with at least one other WTRU (e.g., a plurality of WTRUs), or between at least one other WTRU, and the third information can indicate a computed gradient and an index associated with a training iteration.

[0014] In various embodiments, the method can include or the WTRU can be configured to receive or transmit an evaluation request indicator indicating any of an NBP model index and a Monte Carlo parameter associated with the NBP model; receive or transmit an evaluation ready indicator (ERI); transmit a set of codewords and an index of the codewords; and receive an evaluation completion indicator when a stopping criterion is met.

[0015] In various embodiments, the method can include or the WTRU can be configured to transmit a request to a network element to train a parameter adapter network (PAN), and receive a PAN configuration signal from the network element indicating any of a PAN architecture and parameters, and / or a CSI-RS for estimating a channel.

[0016] In various embodiments, the request to train the PAN can be transmitted with the request to receive a training configuration.

[0017] In various embodiments, the method can include or the WTRU can be configured to select the NBP variant or architecture using a selection criterion based on any one or more of: latency, complexity, reliability, channel conditions, and encoding parameters. BRIEF DESCRIPTION OF DRAWINGS

[0018] A more detailed understanding can be had from the following description, given by way of example in conjunction with the accompanying drawings wherein:

[0019] FIG. 1A is a system diagram illustrating an example communications system;

[0020] FIG. 1B It shows that FIG. 1A A system diagram of an example wireless transmit / receive unit (WTRU) for use within the illustrated communication system;

[0021] FIG. 1C It shows that you can FIG. 1A A system diagram illustrating an exemplary radio access network (RAN) and an exemplary core network (CN) for use within the illustrated communication system;

[0022] FIG. 1D It is a diagram that can be FIG. 1A A system diagram of an additional example RAN and an additional example CN used within the illustrated communication system;

[0023] FIG. 2 shows an example FNSPA decoding architecture

[0024] FIG. 3 Two unfolding examples of NBP architecture based on relaxed RNN are shown;

[0025] FIG. 4 shows an example signal flow diagram according to an example embodiment;

[0026] FIG. 5 An example flow chart illustrating a method for NBP training according to an exemplary embodiment; and

[0027] FIG. 6 An example PAN architecture is shown in accordance with various embodiments. DETAILED DESCRIPTION

[0028] In the following detailed description, many specific details are set forth to provide a thorough understanding of the embodiments and / or examples disclosed herein. However, it will be understood that such embodiments and examples can be put into practice without some or all of the specific details set forth herein. In other cases, well-known methods, processes, components and circuits are not described in detail to avoid confusing the following description. In addition, the embodiments and examples not specifically described herein can replace the embodiments and other examples described, disclosed or otherwise explicitly, implicitly and / or inherently provided (collectively referred to as "providing") herein or be implemented in combination with them. Although various embodiments in which devices, systems, equipment, etc. and / or any of its elements perform operations, processes, algorithms, functions, etc. and / or any part thereof are described and / or claimed herein, it should be understood that any embodiment described and / or claimed herein assumes that any device, system, equipment, etc. and / or any of its elements are configured to perform any operation, process, algorithm, function, etc. and / or any part thereof.

[0029] The methods, apparatus and systems provided herein are well suited for communications involving both wired and wireless networks. With respect to FIGS. 1A-1D An overview of various types of wireless devices and infrastructure is provided, where various elements of the network can utilize, perform, be arranged and / or adapted and / or configured in accordance with the methods, apparatus and systems provided herein, and for the methods, apparatus and systems provided herein.

[0030] FIG. 1A is a system diagram illustrating an example communications system 100 in which one or more disclosed embodiments can be implemented. The communications system 100 can be a multiple access system that provides content, such as voice, data, video, messaging, broadcast, etc., to multiple wireless users. The communications system 100 can enable multiple wireless users to access such content through the sharing of system resources, including wireless bandwidth. For example, the communications systems 100 can employ one or more channel access methods, such as code division multiple access (CDMA), time division multiple access (TDMA), frequency division multiple access (FDMA), orthogonal FDMA (OFDMA), single-carrier FDMA (SC-FDMA), zero-tail (ZT) unique-word (UW) discrete Fourier transform (DFT) spread OFDM (ZT UW DFT-S OFDM), unique-word OFDM (UW-OFDM), resource block-filter OFDM, filter bank multicarrier (FBMC), and the like.

[0031] As FIG. 1AAs shown, the communication system 100 can include wireless transmit / receive units (WTRUs) 102a, 102b, 102c, 102d, a radio access network (RAN) 104 / 113, a core network (CN) 106 / 115, a public switched telephone network (PSTN) 108, the Internet 110, and other networks 112, though it will be appreciated that the disclosed embodiments contemplate any number of WTRUs, base stations, networks, and / or network elements. Each of the WTRUs 102a, 102b, 102c, 102d can be any type of device configured to operate and / or communicate in a wireless environment. By way of example, the WTRUs 102a, 102b, 102c, 102d (any of which can be referred to as a "station" and / or a "STA") can be configured to transmit and / or receive wireless signals and can include (or be) a user equipment (UE), a mobile station, a fixed or mobile subscriber unit, a subscription-based unit, a pager, a cellular telephone, a personal digital assistant (PDA), a smartphone, a laptop, a netbook, a personal computer, a wireless sensor, a hotspot or Mi-Fi device, an Internet of Things (IoT) device, a watch or other wearable device, a head-mounted display (HMD), a vehicle, a drone, a medical device and applications (e.g., remote surgery), an industrial device and applications (e.g., a robot or other wireless devices operating in an industrial and / or an automated processing chain environments), a consumer device, a device operating on a commercial and / or industrial wireless network, etc. Any of the WTRUs 102a, 102b, 102c, and 102d, or any other WTRU referred to herein, can be interchangeably termed a UE.

[0032] The communication system 100 can also include a base station 114a and / or a base station 114b. Each of the base stations 114a, 114b can be any type of device configured to wirelessly interface with at least one of the WTRUs 102a, 102b, 102c, 102d to facilitate access to one or more communication networks, such as the CN 106 / 115, the Internet 110, and / or the networks 112. By way of example, the base stations 114a, 114b can be any of a base transceiver station (BTS), a Node-B, an eNode-B (eNB), a Home Node-B (HNB), a Home eNode-B (HeNB), a gNode-B (gNB), a NR Node-B (NR NB), a site controller, an access point (AP), a wireless router, and so on. While the base stations 114a, 114b are each depicted as a single element, it will be appreciated that the base stations 114a, 114b can include any number of interconnected base stations and / or network elements.

[0033] The base stations 114a can be part of the RAN 104 / 113, which can also include other base stations and / or network elements (not shown), such as a base station controller (BSC), a radio network controller (RNC), a relay node, etc. The base stations 114a and / or the base stations 114b can be configured to transmit and / or receive wireless signals on one or more carrier frequencies, which can be referred to as a cell (not shown). These frequencies can be in the licensed spectrum, the unlicensed spectrum, or a combination of the licensed and unlicensed spectrums. A cell can provide wireless service to a particular geographic area that can be relatively fixed or can change over time as users who are in the cell move. The geographic area of a cell can be further split into cell sectors, which can be at least partially overlapped. For example, the cell associated with a base station 114a can be split into three sectors. Thus, in one embodiment, a base station 114a can include three transceivers, one for each sector of the cell. In an embodiment, the base station 114a can employ multiple-input multiple-output (MIMO) techniques and can utilize multiple transceivers for each or any sector of the cell. For example, beamforming can be used to transmit and / or receive signals in a desired spatial direction.

[0034] The base stations 114a, 114b can communicate with one or more of the WTRUs 102a, 102b, 102c, 102d over an air interface 116, which can be any suitable wireless communication link (e.g., radio frequency (RF), microwave, centimeter wave, micrometer wave, infrared (IR), ultraviolet (UV), visible light, etc.). The air interface 116 can be established using any suitable radio access technology (RAT).

[0035] More specifically, as noted above, the communications system 100 can be a multiple access system and can employ one or more channel access schemes, such as CDMA, TDMA, FDMA, OFDMA, SC-FDMA, and the like. For example, the base station 114a and the WTRUs 102a, 102b, 102c in the RAN 104 / 113 can implement a radio technology such as Universal Mobile Telecommunications System (UMTS) Terrestrial Radio Access (UTRA), which can establish the air interface 116 using wideband CDMA (WCDMA). WCDMA can include communication protocols such as High-Speed Packet Access (HSPA) and / or Evolved HSPA (HSPA+). HSPA can include High-Speed Downlink Packet Access (HSDPA) and / or High-Speed Uplink Packet Access (HSUPA).

[0036] In an embodiment, the base station 114a and the WTRUs 102a, 102b, 102c can implement a radio technology such as Evolved UMTS Terrestrial Radio Access (E-UTRA), which can establish the air interface 116 using Long Term Evolution (LTE) and / or LTE-Advanced (LTE-A) and / or LTE-A Pro.

[0037] In an embodiment, the base station 114a and the WTRUs 102a, 102b, 102c can implement a radio technology such as NR Radio Access, which can establish the air interface 116 using New Radio (NR).

[0038] In an embodiment, the base station 114a and the WTRUs 102a, 102b, 102c can implement multiple radio access technologies. For example, the base station 114a and WTRUs 102a, 102b, 102c can implement LTE wireless access and NR wireless access together, for instance using dual connectivity (DC) principles. Thus, the air interface utilized by WTRUs 102a, 102b, 102c can be characterized by multiple types of wireless access technologies and / or transmissions having varying technologies, e.g., transmitted to / from multiple types of base stations (e.g., a gNB and an eNB) and / or transmitted using multiple types of radio access technologies.

[0039] In an embodiment, the base station 114a and the WTRUs 102a, 102b, 102c can implement a radio technology such as IEEE 802.11 (i.e., Wireless Fidelity (Wi-Fi), IEEE 802.16 (i.e., Worldwide Interoperability for Microwave Access (WiMAX)), CDMA2000, CDMA2000 IX, CDMA2000 EV-DO, Interim Standard 2000 (IS-2000), Interim Standard 95 (IS-95), Interim Standard 856 (IS-856), Global System for Mobile communications (GSM), Enhanced Data rates for GSM Evolution (EDGE), GSM EDGE (GERAN), and / or the like.

[0040] FIG. 1AThe base station 114b in FIG. 1 A can be a wireless router, Home Node-B, Home eNode-B, or access point, for example, and can utilize any suitable RAT for facilitating wireless connectivity access in a localized area, such as a place of business, a home, a vehicle, a campus, an industrial facility, an air corridor (e.g., for use by drones), a roadway, and the like. In one embodiment, the base station 114b and the WTRUs 102c, 102d can implement a radio technology such as IEEE 802.11 to establish a wireless local area network (WLAN). In an embodiment, the base station 114b and the WTRUs 102c, 102d can implement a radio technology such as IEEE 802.15 to establish a wireless personal area network (WPAN). In an embodiment, the base station 114b and the WTRUs 102c, 102d can utilize a cellular-based RAT (e.g., WCDMA, CDMA2000, GSM, LTE, LTE-A Pro, NR, etc.) to establish any of a small cell, pico cell, or femto cell with various levels of area and FIG. 1A As shown, the base station 114b can have a direct connection to the Internet 110. Thus, the base station 114b can not be required to access the Internet 110 via the CN 106 / 115.

[0041] The RAN 104 / 113 can be in communication with the CN 106 / 115, which can be any type of network configured to provide voice, data, applications, and / or voice over internet protocol (VoIP) services to one or more of the WTRUs 102a, 102b, 102c, 102d. The data can have varying quality of service (QoS) requirements, such as differing throughput requirements, latency requirements, error tolerance requirements, reliability requirements, data throughput requirements, mobility requirements, and the like. The CN 106 / 115 can provide call control, billing services, mobile location-based services, pre-paid calling, Internet connectivity, video distribution, etc., and / or perform high-level security functions, such as user authentication. Although not shown in FIG. 1A Although not shown in FIG. 1 A, it will be appreciated that the RAN 104 / 113 and / or the CN 106 / 115 can be in direct or indirect communication with other RANs that employ the same RAT as the RAN 104 / 113 or a different RAT. For example, in addition to being connected to the RAN 104 / 113, which can employ a NR radio technology, the CN 106 / 115 can also be in communication with another RAN (not shown) employing a GSM, UMTS, CDMA2000, WiMAX, E-UTRA, or Wi-Fi radio technology.

[0042] CN 106 / 115 can also serve as a gateway for the WTRUs 102a, 102b, 102c, 102d to access the PSTN 108, the Internet 110, and / or other networks 112. The PSTN 108 can include circuit-switched telephone networks that provide infrastructure for the provision of voice telephony. The Internet 110 can include a global system of interconnected computer networks and devices that use the Transmission Control Protocol (TCP) / Internet Protocol (IP) suite of protocols to facilitate communications and exchange of data between devices. The networks 112 can include other wired or wireless networks that are owned and / or operated by other service providers. For example, the networks 112 can include another CN that is connected to one or more RANs, which can employ the same RAT as the RAN 104 / 114 or a different RAT.

[0043] Some or all of the wireless transmissions and receptions in the communication system 100 can include multi-mode capability (e.g., the wireless transmit and receive units 102a, 102b, 102c, 102d can include multiple transceivers to communicate with different wireless networks over different wireless links). For example, FIG. 1A The WTRU 102c shown in Figure 1C can be configured to communicate with the base station 114a using a cellular-based radio technology and the base station 114b using an IEEE 802 radio technology.

[0044] FIG. 1B is a system diagram illustrating an example WTRU 102. As shown in FIG. 1B The WTRU 102 can include a processor 118, a transceiver 120, a transmit / receive element 122, a speaker / microphone 124, a keypad 126, a display / touchpad 128, non-removable memory 130, removable memory 132, a power source 134, a global positioning system (GPS) chipset 136, and / or other

[0045] The processor 118 can be a general purpose processor, a special purpose processor, a conventional processor, a digital signal processor (DSP), a plurality of microprocessors, one or more microprocessors in association with a DSP core, a controller, a microcontroller, Application Specific Integrated Circuits (ASICs), Field Programmable Gate Array (FPGAs) circuits, any other type of integrated circuit (IC), a state machine, and the like. The processor 118 can perform signal coding, data processing, power control, input / output processing, and / or any other functionality that enables the WTRU 102 to operate in a wireless environment. The processor 118 can be associated with a FIG. 1B The processor 118 and the transceiver 120 are depicted as separate components, but it will be understood that the processor 118 and the transceiver 120 can be integrated together, e.g., as an electronic package or chip.

[0046] The transmit / receive element 122 can be configured to transmit signals to, or receive signals from, a base station (e.g., the base station 114a) over the air interface 116. For example, in one embodiment, the transmit / receive element 122 can be an antenna configured to transmit and / or receive RF signals. In an embodiment, the transmit / receive element 122 can be an emitter / detector configured to transmit and / or receive IR, UV, or visible light signals, for example. In an embodiment, the transmit / receive element 122 can be configured to transmit and / or receive both RF and light signals. It will be appreciated that the transmit / receive element 122 can be configured to transmit and / or receive any combination of wireless signals.

[0047] Although the transmit / receive element 122 is depicted in the WTRU 102 FIG. 1B In one embodiment, the WTRU 102 can include two or more transmit / receive elements 122 (e.g., multiple antennas) to facilitate MIMO technology. In this embodiment, the transmit / receive element 122 can be configured to transmit and / or receive wireless signals, respectively, using multiple antennas.

[0048] The transceiver 120 can be configured to modulate the signals that are to be transmitted by the transmit / receive element 122 and to demodulate the signals that are received by the transmit / receive element 122. As noted above, the WTRU 102 can have multi-mode capabilities. Thus, the transceiver 120 can include multiple transceivers for enabling the WTRU 102 to communicate via multiple RATs, such as NR and IEEE 802.11, for example.

[0049] The processor 118 of the WTRU 102 can be coupled to, and can receive user input data from, the speaker / microphone 124, the keypad 126, and / or the display / touchpad 128 (e.g., a liquid crystal display (LCD) display unit or organic light-emitting diode (OLED) display unit). The processor 118 can also output user data to the speaker / microphone 124, the keypad 126, and / or the display / touchpad 128. In addition, the processor 118 can access information from, and store data in, any type of suitable memory, such as the non-removable memory 130 and / or the removable memory 132. The non-removable memory 130 can include random-access memory (RAM), read-only memory (ROM), a hard disk, or any other type of memory storage device. The removable memory 132 can include a subscriber identity module (SIM) card, a memory stick, a secure digital (SD) memory card, and the like. In other embodiments, the processor 118 can access information from, and store data in, memory that is not physically located on the WTRU 102, such as on a server or a home computer (not shown).

[0050] The processor 118 can receive power from the power source 134 and can be configured to distribute and / or control the power to the other components in the WTRU 102. The power source 134 can be any suitable device for powering the WTRU 102. For example, the power source 134 can include one or more dry cell batteries (e.g., nickel-cadmium, nickel-zinc, nickel metal hydride (NiMH), lithium-ion (Li-ion), and the like), solar cells, fuel cells, and the like.

[0051] The processor 118 can also be coupled to the GPS chipset 136, which can be configured to provide location information (e.g., longitude and latitude) regarding the current location of the WTRU 102. In addition to, or

[0052] The processor 118 can further be coupled to other elements / peripherals 138, which can include one or more software and / or hardware modules / units required to provide additional features, functionality and / or wired or wireless connectivity. For example, the elements / peripherals 138 can include an accelerometer, an e-compass, a satellite transceiver, a digital camera (e.g., for photographs or video), a universal serial bus (USB) port, a vibration device, a television transceiver, a hands free headset, an modules, frequency modulation (FM) radio units, digital music players, media player modules, video game player modules, Internet browser, virtual reality and / or augmented reality (VR / AR) devices, activity trackers, etc. The elements / peripherals 138 can include one or more sensors, which can be one or more of a gyroscope, an accelerometer, a hall effect sensor, a magnetometer, a directional sensor, a proximity sensor, a temperature sensor, a time sensor; a geopositioning sensor; an altimeter, a light sensor, a touch sensor, a magnetometer, a barometer, a gesture sensor, a biometric sensor, and / or a humidity sensor.

[0053] The WTRU 102 can include a full duplex radio for which transmission and reception of some or all signals (e.g., some or all signals associated with particular subframes for both uplink (e.g., for transmission) and downlink (e.g., for reception) can be concurrent and / or simultaneous.

[0054] FIG. 1C is a system diagram illustrating the RAN 104 and the CN 106 according to an embodiment. As described above, the RAN 104 can employ an E-UTRA radio technology to communicate with the WTRUs 102a, 102b, 102c over the air interface 116. The RAN 104 can also be in communication with the CN 106.

[0055] The RAN 104 can include eNode-Bs 160a, 160b, 160c, though it will be appreciated that the RAN 104 can include any number of eNode-Bs while remaining consistent with an embodiment. The eNode-Bs 160a, 160b, 160c can each include one or more transceivers for communicating with the WTRUs 102a, 102b, 102c over the air interface 116. In one embodiment, the eNode-Bs 160a, 160b, 160c can implement MIMO technology. Thus, the eNode-B 160a, for example, can use multiple antennas to transmit wireless signals to, and receive wireless signals from, the WTRU 102a.

[0056] Each of the eNode-Bs 160a, 160b, 160c can be associated with a particular cell (not shown) and can be configured to handle radio resource management decisions, handover decisions, scheduling of users in the UL and / or DL, and the like. As shown, the eNode-Bs 160a, 160b, 160c can communicate with one another over an X2 interface. FIG. 1C

[0057] FIG. 1C ​The CN 106 shown in FIG. 10 can include a mobility management entity (MME) 162, a serving gateway (SGW) 164, and a packet data network (PDN) gateway (PGW) 166. While each of the foregoing elements are depicted as part of the CN 106, it will be appreciated that any of these elements can be owned and / or operated by an entity other than the CN operator.

[0058] MME 162 can be connected to each of the eNode-Bs 160a, 160b, and 160c in the RAN 104 via an S1 interface and can serve as a control node. For example, the MME 162 can be responsible for authenticating users of the WTRUs 102a, 102b, 102c, bearer activation / deactivation, selecting a particular serving gateway during an initial attach of the WTRUs 102a, 102b, 102c, and the like. MME 162 can provide a control plane function for switching between the RAN 104 and other RANs (not shown) that employ other radio technologies, such as GSM and / or WCDMA.

[0059] The SGW 164 can be connected to each of the eNode Bs 160a, 160b, 160c in the RAN 104 via the S1 interface. The SGW 164 can generally route and forward user data packets to / from the WTRUs 102a, 102b, 102c. The SGW 164 can perform other functions, such as anchoring user planes during inter-eNode B handovers, triggering paging when DL data is available for the WTRUs 102a, 102b, 102c, managing and storing contexts of the WTRUs 102a, 102b, 102c, and the like.

[0060] The SGW 164 can be connected to the PGW 166, which can provide the WTRUs 102a, 102b, 102c with access to packet-switched networks, such as the Internet 110, to facilitate communications between the WTRUs 102a, 102b, 102c and IP-enabled devices.

[0061] The CN 106 can facilitate communications with other networks. For example, the CN 106 can provide the WTRUs 102a, 102b, 102c with access to circuit-switched networks, such as the PSTN 108, to facilitate communications between the WTRUs 102a, 102b, 102c and traditional landline communications devices. For example, the CN 106 can include, or can communicate with, an IP gateway (e.g., an IP multimedia subsystem (IMS) server) that serves as an interface between the CN 106 and the PSTN 108. In addition, the CN 106 can provide the WTRUs 102a, 102b, 102c with access to the other networks 112, which can include other wired and / or wireless networks that are owned and / or operated by other service providers.

[0062] Although WTRUs are described in FIGS. 1A-1D representative embodiments as wireless terminals, it is contemplated that in certain representative embodiments such terminals can (e.g., temporarily or permanently) use a wired communication interface to a communication network.

[0063] In representative embodiments, the other network 112 can be a WLAN.

[0064] A WLAN in Infrastructure Basic Service Set (BSS) mode can have an access point (AP) for the BSS and one or more stations (STAs) associated with the AP. The AP can have an access or an interface to a Distribution System (DS) or another type of wired / wireless network that carries traffic in to and / or out of the BSS. Traffic to STAs that originates from outside the BSS can arrive through the AP and can be delivered to the STAs. Traffic originating from STAs to destinations outside the BSS can be sent to the AP to be delivered to respective destinations. Traffic between STAs within the BSS can be sent through the AP, for example, where the source STA can send traffic to the AP and the AP can deliver the traffic to the destination STA. The traffic between STAs within a BSS can be considered and / or referred to as peer-to-peer traffic. Peer-to-peer traffic can be sent between STAs directly using a direct link, e.g., without passing through the AP. In certain representative embodiments, a direct link can use an 802.11e DLS or an 802.11z tunneled DLS (TDLS). A WLAN using an Independent BSS (IBSS) mode can not have an AP, and all STAs in an IBSS can communicate directly with each other. The IBSS mode of communication can sometimes be referred to herein as an "ad-hoc" mode of communication.

[0065] When using an 802.11 ac infrastructure mode of operation or similar, an AP can transmit beacons on a fixed channel (e.g., primary channel). The primary channel can be a fixed width (e.g., 20 MHz wide bandwidth) or dynamically set via signaling. The primary channel can be the operating channel of the BSS and can be used by STAs to establish a connection with the AP. In certain representative embodiments, carrier sense multiple access with collision avoidance (CSMA / CA) with collision avoidance can be implemented, e.g., in 802.11 systems. For CSMA / CA, STAs including the AP (e.g., each STA) can sense the primary channel. If the primary channel is sensed / detected and / or determined to be busy by a particular STA, the particular STA can back off. One STA (e.g., only one station) can transmit in a given BSS at any given time.

[0066] High Throughput (HT) STAs can use 40 MHz wide channels, for example, by combining the primary 20 MHz channel with an adjacent or nonadjacent 20 MHz channel to form a 40 MHz wide channel.

[0067] Very High Throughput (VHT) STAs can support 20 MHz, 40 MHz, 80 MHz, and / or 160 MHz wide channels. 40 MHz and / or 80 MHz channels can be formed by combining adjacent 20 MHz channels. A 160 MHz channel can be formed by combining 8 contiguous 20 MHz channels, or by combining two non-contiguous 80 MHz channels, which can be referred to as an 80+80 configuration. For the 80+80 configuration, after channel encoding, the data can go through a segment parser that can divide the data into two streams. Inverse Fast Fourier Transform (IFFT) processing and time domain processing can be done on each stream separately. The streams can be mapped on to the two 80 MHz channels, and the data can be transmitted by a transmitting STA. At the receiver of the receiving STA, the above described 80+80 configuration operations can be reversed, and the combined data can be sent to the Medium Access Control (MAC) layer, entity, etc.

[0068] Operation modes below 1 GHz are supported by 802.11af and 802.11ah. The channel operating bandwidth and carriers are reduced in 802.11af and 802.11ah relative to those used in 802.11η and 802.1 lac. 802.11af supports 5 MHz, 10 MHz, and 20 MHz bandwidths in the TV White Space (TVWS) spectrum, while 802.11ah supports 1 MHz, 2 MHz, 4 MHz, 8 MHz, and 16 MHz bandwidths using non-TVWS spectrum. According to representative embodiments, 802.11ah can support meter type control / machine type communication (MTC), e.g., MTC devices in a macro coverage area. MTC devices can have certain capabilities, e.g., limited capabilities including support (e.g., only support) for certain and / or limited bandwidth. MTC devices can include a battery with a battery life above a threshold (e.g., to maintain a very long battery life).

[0069] WLAN systems that can support multiple channels and channel bandwidths, e.g., 802.11η, 802.1 lac, 802.11af, and 802.11ah, include a channel that can be designated as a primary channel. The primary channel can have a bandwidth equal to the maximum common operating bandwidth supported by all STAs in the BSS. The bandwidth of the primary channel can be set and / or limited by a STA among all STAs operating in the BSS that supports the minimum bandwidth operating mode. In the example of 802.11ah, for a STA (e.g., a MTC type device) that supports (e.g., only supports) a 1 MHz mode, the primary channel can be 1 MHz wide, even though other STAs in the AP and BSS support 2 MHz, 4 MHz, 8 MHz, 16 MHz, and / or other channel bandwidth operating modes. Carrier sensing and / or network allocation vector (NAV) settings can depend on the status of the primary channel. If the primary channel is busy, e.g., due to a STA (that only supports a 1 MHz operating mode) transmitting to the AP, then the entire available frequency band can be considered busy, even though most of the frequency band remains idle and available.

[0070] In the United States, the available frequency band that can be used by 802.11ah is from 902 MHz to 928 MHz. In Korea, the available frequency band is from 917.5 MHz to 923.5 MHz. In Japan, the available frequency band is from 916.5 MHz to 927.5 MHz. Depending on the country code, the total bandwidth available for 802.11ah is 6 MHz to 26 MHz.

[0071] FIG. 1Dis a system diagram illustrating the RAN 113 and the CN 115 according to an embodiment. As described above, the RAN 113 can utilize NR radio technology to communicate with the WTRUs 102a, 102b, 102c over the air interface 116. The RAN 113 can also be in communication with the CN 115.

[0072] The RAN 113 can include gNBs 180a, 180b, 180c, although the RAN 113 can include any number of gNBs while remaining consistent with an embodiment. The gNBs 180a, 180b, 180c can each include one or more transceivers for communicating with the WTRUs 102a, 102b, 102c over the air interface 116. In one embodiment, the gNBs 180a, 180b, 180c can implement MIMO technology. For example, gNBs 180a, 180b can utilize beamforming to transmit signals to and / or receive signals from the WTRUs 102a, 102b, 102c. Thus, the gNB 180a, for example, can use multiple antennas to transmit wireless signals to, and / or receive wireless signals from, the WTRU 102a. In an embodiment, the gNBs 180a, 180b, 180c can implement carrier aggregation technology. For example, the gNB 180a can transmit multiple component carriers (not shown) to the WTRU 102a. A subset of these component carriers can be on unlicensed spectrum while the remaining component carriers can be on licensed spectrum. In an embodiment, the gNBs 180a, 180b, 180c can implement Coordinated Multi-Point (CoMP) technology. For example, WTRU 102a can receive coordinated transmissions from gNB 180a and gNB 180b (and / or gNB 180c).

[0073] The WTRUs 102a, 102b, 102c can communicate with gNBs 180a, 180b, 180c using transmissions associated with an extensible numerology. For example, OFDM symbol spacing and / or OFDM subcarrier spacing can vary for different transmissions, different cells, and / or different portions of the wireless transmission spectrum. The WTRUs 102a, 102b, 102c can communicate with gNBs 180a, 180b, 180c using subframes or subframes or transmission time intervals (TTIs) of varying lengths, e.g., including different numbers of OFDM symbols and / or lasting varying lengths of absolute time.

[0074] The gNBs 180a, 180b, 180c can be configured to communicate with the WTRUs 102a, 102b, 102c in a standalone configuration and / or a non-standalone configuration. In the standalone configuration, the WTRUs 102a, 102b, 102c can communicate with the gNBs 180a, 180b, 180c without also accessing other RANs (e.g., eNode-Bs 160a, 160b, 160c). In the standalone configuration, the WTRUs 102a, 102b, 102c can utilize one or more of the gNBs 180a, 180b, 180c as a mobility anchor point. In the standalone configuration, the WTRUs 102a, 102b, 102c can use signals

[0075] Each of the gNBs 180a, 180b, 180c can be associated with a particular cell (not shown) and can be configured to handle radio resource management decisions, handover decisions, scheduling of users in the UL and / or DL, support of network slicing, dual connectivity, interworking between NR and E-UTRA, routing of user plane data towards User Plane Function (UPF) 184a, 184b, routing of control plane information towards Access and Mobility Management Function (AMF) 182a, 182b, and the like. As shown, the gNBs 180a, 180b, 180c can communicate with one another over an Xn interface. FIG. 1D As shown, the gNBs 180a, 180b, 180c can be in communication with the AN 180a, 180b, 180c over an N2 interface.

[0076] FIG. 1DThe CN 115, as shown in FIG. 10B, can include at least one AMF 182a, 182b, at least one UPF 184a, 184b, at least one Session Management Function (SMF) 183a, 183b, and at least one Data Network (DN) 185a, 185b. While each of the foregoing elements are depicted as part of the CN 115, it will be appreciated that any of these elements can be owned and / or operated by an entity other than the CN operator.

[0077] The AMF 182a, 182b can be connected to one or more of the gNBs 180a, 180b, 180c in the RAN 113 via an N2 interface and can serve as a control node. For example, the AMF 182a, 182b can be responsible for authenticating the WTRUs 102a, 102b, 102c, support for network slicing (e.g., handling of different protocol data unit (PDU) sessions with different requirements), selection of a particular SMF 183a, 183b, management of the WTRU 102a, 102b, 102c’ s registration area, termination of NAS signaling, mobility management, and the like. The AMF 162 can utilize network slicing, for example, to customize CN support for the WTRU 102a, 102b, 102c based on the type of service being utilized, e.g., depend on whether the WTRU 102a, 102b, 102c is engaged in Ultra-Reliable Low-Latency (URLLC) access, eMBB access, MTC access, and / or the like. Different network slices can be established for different use cases, such as services relying on URLLC access, services relying on eMBB access, services relying on MTC access, and / or the like. The AMF 162 can provide control plane functionality for 5G NR and non-3GPP access network technologies, e.g., Wi-Fi.

[0078] The SMF 183a, 183b can be connected to AMF 182a, 182b in the CN 115 via an N11 interface. The SMF 183a, 183b can also be connected to the UPF 184a, 184b in the CN 115 via an N4 interface. The SMF 183a, 183b can select and control the UPF 184a, 184b and configure the routes for traffic through the UPF 184a, 184b. The SMF 183a, 183b can perform other functions, such as managing and allocating IP address, managing PDU sessions, controlling policy enforcement and QoS, providing downlink data notifications, and the like. A PDU session type can be IP-based, non-IP based, Ethernet-based, and the like.

[0079] The UPF 184a, 184b can be connected to one or more of the gNBs 180a, 180b, 180c in the RAN 113 via an N3 interface, which can provide the WTRUs 102a, 102b, 102c with access to packet- switched networks, such as the Internet 110, to facilitate communications between the WTRUs 102a, 102b, 102c and IP-enabled devices. The UPF 184, 184b can perform other functions, such as routing and forwarding packets, enforcing user plane policies, supporting multi-homed PDU sessions, handling user plane QoS, buffering of downlink packets, providing mobility anchoring, and the like.

[0080] The CN 115 can facilitate communications with other networks. For example, the CN 115 can include, or can communicate with, an IP gateway for facilitating communications between the CN 115 and the PSTN 108. Additionally, the CN 115 can provide the WTRUs 102a, 102b, 102c with access to the other networks, which can include other wired and / or wireless networks that are owned and / or operated by other service providers. In one embodiment, the WTRUs 102a, 102b, 102c can be connected to a local DN 185a, 185b through the UPF 184a, 184b via the N3 interface and an N6 interface between the UPF 184a, 184b and the DN 185a, 185b.

[0081] In view of FIGS. 1A-1D and FIGS. 1A-1D In view of the respective descriptions above, one or more or all of the functions described herein with regard to any of the WTRUs 102a-102d, base stations 114a-114b, eNode-Bs 160a-160c, MME 162, SGW 164, PGW 166, gNBs 180a-180c, AMF 182a-182b, UPF 184a-184b, SMF 183a-183b, DN 185a-185b, and / or any other element / device described herein can be performed by one or more emulation elements / devices (not shown). The emulation devices can be one or more devices configured to emulate one or more or all of the functions described herein. For example, the emulation devices can be used to test other devices and / or to simulate a network and / or WTRU functionality.

[0082] The one or more emulation devices can perform one or more, including all, functions while not being implemented / deployed as part of a wired and / or wireless communication network. For example, one or more emulation devices can perform one or more or all functions in a testing scenario in a testing laboratory and / or a non-deployed (e.g., testing) wired and / or wireless communication network in order to implement testing of one or more components. The one or more emulation devices can be test equipment. The emulation devices can transmit and / or receive data using direct RF coupling and / or wireless communication via RF circuitry (e.g., which can include one or more antennas).

[0083] The one or more emulation devices can perform one or more, including all, functions while not being implemented / deployed as part of a wired and / or wireless communication network. For example, one or more emulation devices can perform one or more or all functions in a testing scenario in a testing laboratory and / or a non-deployed (e.g., testing) wired and / or wireless communication network in order to implement testing of one or more components. The one or more emulation devices can be test equipment. The emulation devices can transmit and / or receive data using direct RF coupling and / or wireless communication via RF circuitry (e.g., which can include one or more antennas).

[0084] Embodiments disclosed herein are representative and do not limit applicability of apparatuses, procedures, functions, and / or methods to any particular wireless technology, any particular communication technology, and / or other technologies. The term network in this disclosure can generally refer to one or more base stations or gNBs or other network entities, which in turn can be associated with one or more transmission / reception points (TRPs), or any other node in a radio access network.

[0085] Note that the terms “serving base station,” “base station,” “gNB,” (collectively “gNBs”) can be used interchangeably throughout the example embodiments described herein to designate any network element, such as for example, a network element acting as a serving base station. The embodiments described herein are not limited to gNBs and are applicable to any other type of base station.

[0086] Certain embodiments generally relate to channel decoding variable selection and feedback, deep neural network (DNN) architecture selection and feedback, NBP decoder online training, NBP online evaluation, and / or signal-to-noise ratio (SNR) adaptive training. For example, some embodiments can provide methods for performing online training of a neural belief propagation (NBP) decoder, including user-centric selection for different variants and DNN architectures, and evaluation of machine learning (ML) based decoding models, as well as SNR adaptive training, e.g., for a UE configured to support various NBP decoders and architectures for decoding different families of channel codes, such as low-density parity-check (LDPC) codes, polar codes, Reed-Muller codes, and Bose-Chaudhuri-Hocquenghem (BCH) codes.

[0087] Neural belief propagation (NBP) decoding is a promising decoding scheme adopted in future 3GPP releases and different use cases. NBP enables decoding of different families of error correcting codes, including LDPC codes, polar codes, RM codes, and BCH codes. Since DNNs are used to represent the graph structure of the decoders that can be trained, the hardware implementation complexity is very reasonable and can be parallelizable. DNNs provide flexibility in selecting various architectures and training paradigms in addition to the simplified variants of the BP-SP decoding algorithm.

[0088] These features can enable a UE to select from different architectures and variants depending on its available resources and channel conditions. This flexibility helps to optimize decoding performance as well as latency and complexity depending on the UE, while also allowing for continuous learning on different variants and architectures.

[0089] If C(n, k) is a block code with dimension (or information block length) k and code length n, the code C is characterized by a parity check matrix H, which has M rows and n columns, where M ≥ n - k. The Tanner graph of the code is its graphical representation, which is a bipartite graph with M check nodes and n variable nodes. The edges between a check node c and a variable node v are denoted by e = (c, v). th CN connects to v th VN; L max denotes the maximum number of decoding iterations, and x i,e＝(v,c) denotes the message passed between i th th decoding iteration at c th CN and v th VN.

[0090] NBP decoding is the result of a hybrid of DNNs and belief propagation (BP) decoding. The core concept relies on unfolding the graphical representation of a linear block code (Tanner graph) over all decoding iterations and interpreting the resulting graph as a deep neural network with edges connecting VNs to CNs. The DNN, which can be alternatively referred to as NBP here, can be properly trained using an optimizer, where the edges are iteratively trained to improve decoding performance (BER, BLER).

[0091] One of the main features of NBP decoders is that they provide flexibility in choosing between different architectures related to DNNs, while implementing different methods for training and including other activation functions. Moreover, NBP decoding allows to compensate for the negative effects of short cycles, which are known to have a significant limitation on the decoding performance of traditional message passing decoding algorithms, especially for short codes that are unavoidable for short cycles.

[0092] NBP decoders can be applied to different families of error correcting codes, including LDPC codes, polar codes, RM codes, and BCH codes. Moreover, if properly learned, they can decode encodings with fewer decoding iterations compared to basic BP decoders, while enhancing performance and allowing for parallelizable hardware implementation and processing.

[0093] In terms of complexity, NBP can be applied to different simplified variants of the original BP-SP algorithm, which provides a good solution for low complexity and low latency decoding (e.g., for URLLC and mMTC use cases), while reducing error rates. This is very beneficial, for example, for high reliability and high speed communications that require short block lengths of information.

[0094] BP decoding algorithms can provide very good performance (especially for long block lengths), while maintaining low latency and low complexity. BP can decode any linear block code, however, its main advantage is to operate on sparse parity check matrices in order to provide low latency and low error propagation, as the number of short cycles is small.

[0095] The BP decoding algorithm is an iterative message passing algorithm that operates on the Tanner graph representation of an encoding. It mainly consists of three dependent steps, besides an initialization step, which are check node (CN) update, variable node (VN) update, and marginalization.

[0096] The CN update can include computing the reliability message (LLR) passed from the CN to the VN by computing the sum of extrinsic information provided from the VNs participating in each CN, excluding the contribution of the CN itself. The CN update is referred to as a horizontal step, where the rows of the PCM are considered.

[0097] Similarly, the VN update can compute the extrinsic reliability of the considered CNs that the VN participates in, and is referred to as a vertical step.

[0098] The marginalization step may represent the final operation of the iterative decoding process and may output an estimate of the LLRs of the bits being decoded.

[0099] The BP decoder may have a stopping criterion, which may be the syndrome weight, i.e., when the hard decision of the received sample is a valid codeword, then the syndrome weight is zero, which means that all parity check equations are satisfied. Another stopping criterion may be reaching the maximum number of decoding iterations L max .

[0100] The main variant of the BP decoder is BP-SPA, which guarantees robust performance across the entire SNR range. However, a limitation of this variant is the slightly higher complexity CN update step, which requires hyperbolic arithmetic operations. In some use cases, this limitation is undesirable, as some communication systems require very low latency, and some UEs are energy-constrained and have low memory capacity. To simplify CN updates, other variants such as MS, NMS, OMS, and NOMS exist, and they represent a good trade-off for the performance / complexity trade-off.

[0101] The CN update of BP-SPA can be given by:

[0102]

[0103] VN updates can take the following forms:

[0104] x i，e＝(v，c) =l v +∑ e′＝(v,c′)c′≠c x i-1，e′

[0105] For v th VN, the marginalization step includes the following calculations:

[0106] o v =l v +∑ e′＝(v，c′) x i-1，e′

[0107] where l v is with v th Bit-correlated LLR input.

[0108] NBP is a most attractive solution for decoding all linear block codes with low complexity and low latency. Making the decoding block data-driven and further enhancing the performance is also a promising approach.

[0109] NBP can include considering an expanded version of the Tanner graph on all decoding iterations and learning the edges connecting variable nodes to check nodes. The motivation for using NBP to decode linear block codes lies in the fact that NBP allows to compensate the negative effects of short cycles, which are known to be one of the main causes of performance degradation in BP. If the edges are learned optimally, the messages passed from the edges resulting from short cycles are attenuated, which pushes the decoder to minimize the error propagation effects caused by the presence of short cycles.

[0110] Similarly to the basic BP, NBP can be applied to different variants of BP, which enables to consider these simplified check node updates in order to reduce complexity and latency, while improving the performance compared to the basic BP-SPA. Moreover, NBP can also be used to decode PR-LDPC codes in order to improve their performance.

[0111] NBP can be represented by a DNN, where the input layer is fed by the LLR initialization, which means that the number of processing units (neurons) in the input layer is equal to the code length n. The DNN structure is composed of several hidden layers, where each expansion of hidden layers represents a VN update layer followed by a CN update layer. Depending on the architecture of NBP, these layers can be followed by a marginalization layer, in which case the DNN can not be fully processed during the inference phase, since the marginalization layer outputs the LLR estimates that enable to compute the companion-wise weights used as a stopping criterion. The output layer of the NBP model includes n LLR estimates activated using a Sigmoid function. In this case, NBP has an RNN structure composed of L max expansions. In another case, there is only one marginalization layer as output layer, and in this case the number of hidden layers is equal to 2L max , and the architecture is feed-forward (FF). Generally, the DNN graph architecture of NBP is related to the Tanner graph, and the layers are not fully connected. The neurons of NBP are the nodes of the Tanner graph (CN and VN), and the edges are the connections between these nodes (non-zero elements of the PCM associated to the encoding).

[0112] The edges of NBP can be trained on all iterations and have different weights, in which case the NBP architecture is FF. When the weights are linked in time, i.e. the edges share the same weight during all iterations, then the NBP architecture is RNN and includes a repetition of the same expansion (VN layer -> CN layer -> marginalization layer). In another option, the weights can be linked in space and time, which means that all edges share the same weight between all decoding iterations. This simplified architecture is called simple scaling (SS) and helps to significantly reduce the storage requirements, training and decoding complexity.

[0113] The computations involved in each hidden layer type of NBP are shown below. Also, different CN updates associated with different BP variants are denoted. Note that these updates correspond to the FF architecture of different variants FNSPA, FNMS, and FNOMS.

[0114] • VN layer:

[0115]

[0116] • CN layer:

[0117] • FNSPA:

[0118]

[0119] • FNMS:

[0120]

[0121] • FNOMs:

[0122]

[0123] • Marginalization layer:

[0124]

[0125] where is the sigmoid activation function.

[0126] The inner weights w i，v and the marginalization weights w 2L+1，v have been shown to have negligible impact on performance and can thus be discarded; in contrast to the weights used in the VN and CN layers of the simplified variants, where in FNMS the trainable weights are normalization factors and in FNOMS are offset parameters. Note that when the trainable weights are equal to 1, then the algorithm is similar to the basic BP. Thereafter, if learned correctly, the NBP decoder is guaranteed to perform better than the regular decoder if the weights are initialized to 1 before training. In some examples, the weights can be or can correspond to previously trained weights. For example, if the NBP has already been trained, then the weights can be initialized by their learned values.

[0127] For the RNN architecture of NBP, i.e., variants RNSPA, RNMS, and RNOMS, the trainable weights can be simplified to w ie＝(v，c) = w e＝(v，c) . For the SS scheme, the weights can share the same value for all edges and decoding iterations, so w i，e＝(v，c) = w.

[0128] FIG. 2 and FIG. 3FF and RNN based architectures for NBP are shown in Fig. 1. FIG. 2 An example FNSPA decoding architecture with three decoding iterations for n = 15 is shown. As FIG. 2 shown, a Tanner graph representing the unfolding over the decoding iterations (5 iterations in this case) is shown. The input nodes are the n = 15 received symbols (channel output). The green nodes in the first hidden layer correspond to both CN and VN updates, as the messages are initialized to zero, so only CN updates are computed. The blue nodes correspond to VNs, while the red nodes are CNs. The edges connecting the node layers are the positions of the non-zero terms in the PCM, based on which the reliability messages are computed and passed. The output of the DNN is the result of the marginalization layer, which outputs n estimates of the LLRs corresponding to each encoded bit.

[0129] To further improve the performance of NBP by ML techniques, another trainable parameter, the so-called relaxation or dumping parameter γ, can be added, and the resulting decoder is relaxed. The main benefit of the relaxation factor is to help speed up the convergence rate and improve performance by applying an exponentially weighted moving average to combine the messages sent at i - 1th iteration with the original messages computed in the ith iteration. This continuous relaxation process yields the subsequent filtered messages where γ ∈ [0, 1] is the relaxation factor. The decoder is less relaxed for γ → 0, and more relaxed for γ → 1. The decoder has no relaxation at γ = 0. FIG. 3 Two examples of the unfolding of the NBP architecture based on relaxed RNNs are shown.

[0130] Training the NBP decoder can incorporate different choices of the objective function. The most common one is the binary cross-entropy (BCE) multi-loss function, which has the advantage of updating the gradients after each decoding iteration, resulting in performance improvement and faster training convergence, as opposed to the regular BCE. This multi-loss BCE function can take the following form:

[0131]

[0132] where o v，i is the output of the neural network for the v th component of the codeword used for transmission at the i th th decoding iteration, while y v is the actually transmitted codeword symbol (label). Note that due to the symmetry property, the decoder can be trained using the all-zero codeword transmitted, in which case y v = 0 for all 1 ≤ v ≤ n.

[0133] Other loss functions can also be considered in the NBP scheme, such as a soft bit error rate loss function, or a soft- companion weighted loss function, which allows the training to be unsupervised. Moreover, these functions can be combined into a general loss function.

[0134] Generally, artificial intelligence (AI) can be broadly defined as behavior exhibited by machines. Such behavior can, for example, mimic cognitive functions to perceive, reason, adapt, and / or act.

[0135] Machine learning (ML) can refer to a category of algorithms that solve problems based on learning through experience ("data") rather than being explicitly programmed ("configuring a rule set"). Machine learning can be considered a subset of AI. Different machine learning paradigms can be envisioned based on the nature of the data or feedback available to the learning algorithm. For example, supervised learning methods can involve learning a function that maps inputs to outputs based on labeled training examples, where each training example can be a pair consisting of an input and a corresponding output. For example, unsupervised learning methods can involve detecting patterns in data for which there are no pre-existing labels. For example, reinforcement learning methods can involve executing sequences of actions in an environment to maximize cumulative rewards. In some examples, machine learning algorithms can be applied using a combination or interpolation of the above methods. For example, semi-supervised learning methods can use a combination of a small amount of labeled data with a large amount of unlabeled data during training. In this regard, semi-supervised learning falls between unsupervised learning (no labeled training data) and supervised learning (only labeled training data).

[0136] Deep learning (DL) refers to a category of machine learning algorithms that employ artificial neural networks (e.g., DNNs) loosely inspired by biological systems. A deep neural network (DNN) is a special class of machine learning models inspired by the human brain, in which inputs are linearly transformed and passed through non-linear activation functions multiple times. DNNs typically include multiple layers, where each layer contains linear transformations and a given non-linear activation function. DNNs can be trained using training data via backpropagation algorithms. Recently, DNNs have shown state-of-the-art performance in various domains, e.g., speech, vision, natural language, etc., and are used in various machine learning settings including supervised, unsupervised, and semi-supervised. The term AI-based method / process can refer to behavior and / or compliance with requirements achieved by learning based on data without explicit configuration of a sequence of actions. Such methods can enable learning of complex behaviors that can be difficult to specify and / or implement when using traditional methods.

[0137] As mentioned above, the belief propagation decoder is a channel decoding algorithm that can be used in different communication systems for decoding LDPC codes and their different variants. BP provides good performance while requiring a relatively low decoding latency.

[0138] One challenging problem involves decoding other code families such as polar codes, RM codes, and BCH codes with BP algorithms while getting reasonable performance. Moreover, LDPC codes show limitations in the short block length range due to the presence of short cycles in their dual structure, similarly to other linear block codes, which significantly affect the error rate performance. One such use case that requires short block length communication includes, but is not limited to, 5G URLLC and mMTC use cases, vehicular communications, and other standards in which encodings such as polar codes or BCH codes need to be decoded by low complexity and latency algorithms while maintaining their performance robustness. Another problem is that the current channel coding schemes in 5G NR are static and not data-driven, and they do not take into account channel conditions or communication scenarios or use cases. Moreover, user-centric channel coding is another feature that can lead to performance improvements in the next releases.

[0139] One particular example is the decoding of polar codes. Polar codes are capacity achieving codes under CRC-aided serially concatenated list decoding (CA-SCL), and they have been adopted in 5G NR eMBB for control channels. However, CA-SCL decoding requires very high latency and implementation complexity, which is not necessarily supported by some UE devices and use cases. Also, BCH codes have proven their capability to approach the performance in the short block length range when decoded with an ordered statistics decoder (OSD). Similarly, OSD is a very complex decoder, and although many variants aiming to reduce its complexity have been proposed in the literature, this algorithm is potentially not recommended for many use cases that require high-speed communications.

[0140] One promising solution for optimizing the performance / latency / complexity trade-off for short block length communications and different encodings is to use neural BP decoders that have the feature of implementing data-driven decoding based on various channel types and conditions while learning the parameters of the decoders that allow to compensate the negative effects of short cycles, while maintaining the low latency properties of these decoders at the cost of performance improvements. NBP decoding has shown attractive performance for various code families and has the feature of adapting to different channel conditions, as well as different degrees of freedom for further optimization. NBP decoding can eventually be the most suitable decoding scheme for PR-LDPC BG that has been adopted in 5G NR eMBB, and it will likely be considered for other 5G use cases and 6G communications as the most probable candidate channel coding scheme.

[0141] Representative embodiments discussed herein can address several problems. For example, some problems addressed by some example embodiments can include at least the following: how to overcome the negative impact of short periods on the performance of current channel coding schemes (5G NR) and, in general, for different codes characterized by PCM and decoded by message passing algorithms; how to further improve the performance (BLER) of 5G NR channel coding, especially for short block length transmissions that are performance limited; how to reduce the decoding latency; how to enable a dynamic data driven channel coding mechanism instead of a static approach; how to enable user-centric channel decoding; and how to enable message exchange between UE and gNB to support online NBP training and performance monitoring.

[0142] In example embodiments discussed herein, methods are described for addressing the use of different variants and architectures of NBP decoders for decoding arbitrary classes of codes. For example, certain embodiments include methods for UE pre-configuration, online NBP joint and distributed training, enablers and / or procedures for SNR adaptive training of UE to parameter adapter networks, methods for selecting NBP variants and / or architectures according to various user-centric metrics, and / or methods for evaluating and updating weights associated with trained decoders.

[0143] It should be noted that, in certain embodiments described herein, NBP variants can be interchangeably referred to as NBP architectures. Additionally, in some embodiments, NBP variants or NBP architectures can alternatively refer to or include NBP models. Thus, as described herein, NBP variants or architectures can be referred to as NBP models.

[0144] Certain embodiments can include methods, apparatuses, systems, etc. that can be directed to individual and / or joint training and evaluation of NBP decoders.

[0145] In various embodiments, the methods for training and / or evaluating the NBP decoder can be implemented in a WTRU, and / or the methods used in conjunction with training and / or evaluating the NBP decoder can be implemented in a WTRU. According to some embodiments, the WTRU can perform training individually, jointly / separately with a gNB, and / or distributedly / collaboratively with other WTRUs. In one embodiment, the method can include any of the following: receiving a training indicator (TI) from a gNB; sending a training configuration request (TCR) to the gNB along with FEC code parameters; receiving a training configuration (TC) from the gNB; sending a training ready indicator (TRI) to the gNB; subsequently receiving a data set from the gNB; sending a training progress indicator (TPI) to the gNB, for example, after each training round, after every M rounds, and / or after training is completed; and communicating a training completion indicator (TCI), for example, after completing all training rounds and / or if the loss function of the model starts to diverge beyond a predefined threshold T. In various embodiments, the aforementioned messaging steps can be performed bi-directionally (e.g., to / from the WTRU and gNB) if the training is joint and / or separate.

[0146] In various embodiments, the methods for updating the NBP model and performance table and / or the methods used in conjunction with updating the NBP model and the performance table can be implemented in a WTRU. According to some example embodiments, the method can include receiving a BLER reference signal (BRS) from a gNB, which can contain pairs of SNR / BLER values and / or NBP model indices. The method can further include comparing the newly updated BLER performance with a stored BLER table, deciding whether the weights of the NBP model should be updated, and / or sending an updated model indicator (UMI) to the gNB.

[0147] Certain embodiments can include methods, apparatuses, systems, and the like, which can be directed to distributed training and evaluation of NBP decoders. For example, for distributed training, one or more (e.g., several or multiple) WTRUs can participate, and a gNB can initiate training by, for example, periodically, aperiodically, and / or triggered by a WTRU, communicating and / or broadcasting a training indicator (TI). In one embodiment, the method can include receiving the TI by each WTRU participating in the distributed training / evaluation. The method can include the WTRU communicating a TCR to the gNB, receiving a training configuration signal (TCS) from the gNB, and / or communicating the computed gradients and training round indices to the gNB. In one embodiment, the method can further include communicating a training completion indicator (TCI) to the gNB after completing the training process.

[0148] For evaluation of the NBP, the method can include the WTRU transmitting and / or receiving an evaluation request indicator (EQI) that contains or indicates, for example, an NBP model index and / or a Monte Carlo parameter. Alternatively or additionally, the EQI can be broadcast to one or more (e.g., several or multiple) WTRUs if the evaluation is distributed. In one embodiment, the method can include the WTRU receiving an evaluation ready indicator (ERI) from the gNB and / or transmitting the ERI to the gNB in case of distributed evaluation. According to some embodiments, the method can include the WTRU transmitting a set of codewords and their indices and / or receiving an evaluation complete indicator (ECI) from the gNB when a stopping criterion is met.

[0149] Certain embodiments can include methods, apparatus, systems, etc. that involve SNR adaptive training. In various implementations, methods for SNR adaptive training and / or methods used in connection with SNR adaptive training can be implemented in a WTRU. According to some implementations, the method can include the WTRU training a PAN model jointly with an NBP model or separately (e.g., serially) from the NBP model. In case of joint training using the NBP model, the method can include the WTRU transmitting a parameter adapter training request (PATR) to the gNB along with the TCR. In case of training the PAN model separately (e.g., serially) from the NBP model, the method can include transmitting the PATR to the gNB and receiving a PAN configuration signal (PCS) from the gNB. The PCS can contain or indicate a PAN architecture and parameters (e.g., can be implicit or explicit) and / or a CSI-RS used for estimating the channel.

[0150] Additionally, certain embodiments can include methods, apparatus, systems, etc. that can be directed to inference of the NBP model.

[0151] Some example embodiments can include methods for configuring the NBP model. In various embodiments, the UE can be configured to support different NBP architectures and variants, such as (but not limited to) FNSPA, RNSPA, FNNMS, RNNMS, NSPA-SS, and / or NNMS-SS. Note that this is not to be an exhaustive list of applicable architectures, as applicable architectures can be broader than those listed above; however, these algorithms are provided as examples to consider different use cases and address any requirements.

[0152] Given an encoding with parity check matrix, the objective can include implementing NBP decoding by training edges of the associated Tanner graph connecting VN to CN, i.e., non-zero elements of the PCM. The PCM can characterize a PR-LDPC code, e.g., a base graph (BG) used in 5G NR eMBB in a data channel, or a single PCM corresponding to an LDPC code, or an arbitrary PCM associated with a polar code, a BCH code, or an RM code.

[0153] E can be or can represent the number of edges contained in the Tanner graph, or equivalently, the number of non-zero elements in the PCM. Assuming the decoder under consideration is relaxed, the total number of trainable parameters stored by the UE can depend on the NBP architecture and / or variant under consideration. Since the parameters of RNN and SS architectures are linked in time and space / time, these architectures can require fewer trainable parameters. The number of parameters can be independent of the variable (SPA or NMS). The parameters can include weights connecting CN to VN, in addition to one considered as a relaxation factor tied to iteration and edge. Thus, the number of parameters for FF structures (e.g., FNSPA and FNNMS) can equal (E x L) + 1, while for RNN structures (e.g., RNSPA and RNNMS), the number of parameters can equal E + 1. For SS architectures (e.g., NSPA-SS and NNMS-SS), there can be only 2 parameters to train, since all edge weights are linked. Table 1 below summarizes the total number of trainable parameters associated with different NBP decoding architectures.

[0154]

[0155]

[0156] Table 1

[0157] According to some embodiments, the UE can be configured to store a table mapping or associating the parameters to different NBP variants and / or architectures. Note that each variant belonging to the same architecture, although having the same number of weights, can have different trained parameter values. The UE can also be preconfigured with the maximum number of decoding iterations L max indicating the number of hidden layers in the DNN architecture of the NBP.

[0158] To train a BG corresponding to a PR-LDPC code, such as those used in 5G NR (BG1 and BG2), the UE can be configured to store parameters associated with an edge bundle, i.e., the same edge bundle shares the same trained weights, allowing the same weights to be extended to all edges belonging to the same bundle after a lifting operation defined by a lifting factor Z, which is determined during a rate matching procedure. In another option that can be similar to training a single PCM, the UE can be configured to store parameters associated with one PCM that is generated by lifting a BG. This can be very useful to train individual edges in cases where the same code length is frequently used in a particular use case or application. This allows the training to be more targeted to individual edges, to further enhance performance.

[0159] Other configurations that the UE can consider can include ML hyperparameters for the NBP. For example, for each individual decoder, the UE can be preconfigured by an optimizer algorithm to sufficiently update the gradient, including SGD, ADAM, and RMSPROP. Each optimizer can require its own parameters, e.g., ADAM requires momentum and decay factor parameters. Also, the UE can be configured by a learning rate a associated with each optimizer and NBP decoder. Other hyperparameters can include batch size, number of training epochs, stopping criteria, and objective function to be combined. Different choices of objective functions are described below or elsewhere herein.

[0160] Some example embodiments can include methods for UE online training and / or evaluation of NBP decoders. In one embodiment, the UE can be configured to perform online NBP training, e.g., after being preconfigured. In this example, the UE can receive a dataset from a gNB, which can then transmit the dataset over different channel conditions.

[0161] FIG. 4An example signal flow diagram is shown in accordance with one example embodiment. In various embodiments, as shown at 405, a gNB can periodically transmit a training request to a UE using a training indicator (TI) over a PDCCH. The TI can be sent semi-periodically after a predefined timing configured by the network or periodically. Further, in another example, the TI can contain an indication on the number of OFDM symbols, slots, and / or subframes before the training starts. In the case where the UE is ready (e.g., prepared) to conduct the training, as shown at 410, the UE can send a training configuration request (TCR) to the gNB, e.g., over a PUCCH, indicating or requesting to receive information on which NBP architecture to train and / or on hyperparameter updates related to that particular NBP decoder variant or architecture. Otherwise, if the UE is not able to participate in the training or cannot meet the indicated timing, the UE can respond with a negative acknowledgement indicator, e.g., via a PUCCH, or the UE can simply not respond, thereby implicitly indicating (e.g., the gNB knows) to the gNB that the UE is not ready.

[0162] After sending the TCR, as shown at 420, the UE can receive training configuration information (e.g., a training configuration signal (TCS)) over a PDSCH, which contains the index of the NBP variant to train and any hyperparameter updates. For each NBP model, the hyperparameters are considered known to the UE by means of configuration data containing a lookup table mapping or associating each NBP variant with its hyperparameters and weights. The UE can also receive in the TCS an indication of which BG to train (e.g., BG1 or BG2 in 5G NR eMBB) and / or other indications related to channel coding parameters, e.g., including code length N and code rate r. In one embodiment, after receiving all parameters, as shown at 425, the UE can signal to the gNB that it is ready to start the training, e.g., by sending a training ready indicator (TRI) over a PUCCH. The TRI can be useful to ensure that the UE refuses to perform the training in case of limited resources or battery life of the UE or in case of very poor channel link with the gNB, e.g., in case of underground or fast mobility. As shown at 430, the gNB can then proceed by progressively transmitting all-zero code words. In one example embodiment, after each training round, the UE can signal or indicate the cost function value to the gNB, e.g., using a training progress indicator (TPI) over a PUCCH, as shown at 435a and 435b. In another example embodiment, the TPI can be fed back periodically after a set of M iterations, e.g., one or more iterations. In yet another example embodiment, the TPI can be sent after the training is completed.

[0163] To provide flexibility, different options can be provided with respect to the TPI. For example, in one example, the TPI can contain the cost function value. In another example, the TPI can contain an indication of whether the cost function is increasing or decreasing, in which case only a 1-bit indicator can be used (e.g., 1 can indicate increasing, and 0 can indicate decreasing). The TPI can be used to monitor the training progress and decide whether the model is converging or diverging. If the model starts to diverge, the gNB can decide to stop the training, e.g., if the cost function reaches a predefined threshold T (included in the PE configuration). Other alternatives for the TPI can be considered depending on the implementation.

[0164] In various embodiments, the UE and the gNB can perform the training jointly or separately (partially aligned or not aligned). In this case, the same signaling can be employed, however, in a bidirectional manner. In this example of joint or separate training, the TI to initiate the training can also be sent by the UE to the gNB to request joint training. Depending on whether the UE or the gNB initiates the process by the TI, the same procedure can be followed for the respective direction. In this case, the dataset can be exchanged and the TPI can also be exchanged subsequently. As mentioned above, the TPI can be transmitted periodically, semi-periodically, or after processing a certain number of training rounds (e.g., processing all training rounds).

[0165] After processing the training iterations (e.g., this can refer to processing an appropriate number of iterations to consider the training complete or processing all training iterations), the UE can indicate to the gNB that the training is complete, e.g., by transmitting a training completion indicator (TCI) via PUCCH, as shown at 440 in the example of FIG. 4. In another example, the UE can use a divergence threshold as a stopping criterion, where the UE can check whether the cost function value is continuously diverging after a certain number T of training rounds, where T is the divergence threshold (i.e., if the cost function is continuously increasing after T training rounds, the model should stop the training process). In this case, the UE can store the values of the weights obtained at the round before the ML model starts to diverge. FIG. 4

[0166] Further, in the case of joint training, the TCI can be sent by the UE or the gNB, depending on which one has completed the training of the NBP model. When transmitting and no TCI is received, the UE or the gNB can continue to send the training dataset until the TCI is received. After training the model by both the UE and the gNB, the UE can send its weights to the gNB, which is done by averaging all available weights.

[0167] ​Upon or after training is complete, the UE may transmit the newly trained weights to the gNB, e.g., via PUSCH, for use in a performance monitoring (evaluation) phase, as shown at 445. In one example, the transmitted weights may be compressed and encoded using LDPC BG1 to enable error detection and correction.

[0168] FIG. 5 An exemplary flow chart of a method for NBP training according to an exemplary embodiment is shown. In various embodiments, FIG. 5 The example method may be performed or implemented by a UE. However, in some embodiments, FIG. 5 The method may be executed or implemented by other network elements or network nodes.

[0169] like FIG. 5 As shown in the example of FIG, at 505, the UE may be configured with or may receive configuration information that includes (or indicates) a lookup table that maps (i.e., associates) NBP models with parameters. For example, the parameters may include any one or more of weights, objective functions, optimizers, hyperparameters, encoding parameters, etc. At 510, the UE may receive a request to perform training (e.g., a training indicator (TI) implemented via a PDCCH, which may be a 1-bit indicator). As described above, the training request (e.g., TI) may be sent semi-periodically or periodically after a predefined timing configured by the network. In some examples, the TI may include an indication of the number of OFDM symbols, time slots, and / or subframes before the training begins. If or when the UE is ready (e.g., prepared) for training, as shown at 515, the UE may send a request for training configuration (e.g., a training configuration request (TCR)) to the gNB (e.g., via a PUCCH), indicating that the UE is ready to receive information regarding which NBP architecture or variant to train and / or hyperparameter updates related to that particular NBP decoder variant or architecture. The UE may then receive a training configuration (e.g., in a training configuration signal (TCS) via PDSCH). At 520, the UE may send a training readiness indicator (e.g., via PUCCH) indicating that it is ready to train the NBP variant / architecture, and at 525, the UE may receive training data and send one or more periodic training progress indicators (e.g., via PDSCH and PUCCH) after one or more rounds or iterations of training. At 530, after or upon completion of training of the NBP variant or architecture, the UE may send an indication of training completion, such as a training completion indicator (TCI) as described above (e.g., a 1-bit indication via PUCCH). At 535, the UE may send (e.g., via PUSCH) newly trained weights associated with the trained NBP variant or architecture, e.g., for use in evaluating the performance of the trained NBP variant or architecture.

[0170] For example, under the 5G NR eMBB use case, the channel coding scheme for data channels relies on PR-LDPC codes defined by 2 different BGs. 3GPP currently adopts BG1, which targets medium to high information block lengths 500≤K≤8448, and medium to high code rates While BG2 targets shorter to medium information block lengths 40≤K≤2560 and lower to medium code rates Thousands of parity check matrices with different lengths and code rates can be generated from these 2 BGs.

[0171] For PR-LDPC codes, for simplicity and practicality, only the edge bundles are trained, and after extending the BG to a PCM, edges originating from the same bundle share the same trained weight. This is the default training configuration, and the UE can indicate to the gNB to activate this configuration by sending a bundle edge indicator (BEI) signal to the gNB, where one bit equals 1 for the default configuration.

[0172] In some use cases or applications, the UE can subsequently use the same information block length, which means using the same PCM associated with the extended BG. In this case, the UE can request to train all edges of the PCM, which can result in performance enhancement. In this case, the BEI equals 0 since a single PCM associated with a single LDPC code is trained.

[0173] After performing the NBP training as detailed above, the UE and gNB can evaluate the AI / ML decoding model before updating the trained weights. Therefore, monitoring the performance of the NBP requires a single transmission (without HARQ process). Before the evaluation process starts, the UE and gNB can be dynamically configured by a performance table that maps some specific SNR values to the associated BLER performance. In other words, the UE and gNB can receive configuration information that includes or indicates a table that associates certain specific SNR values with associated BLER performance. Thereafter, if the BLER performance of the new model is superior to the previously stored model performance, the model can be easily updated after the evaluation process.

[0174] In various embodiments, the evaluation of the NBP model can use online Monte Carlo simulations to collect enough errors per SNR value and obtain the accurate BLER performance of the model. Monte Carlo simulations of a large set of codewords over different SNR values can require computational resources and energy consumption as the NBP decoder performs many decoding trials on the evaluation set. This can not be sufficient for UE devices as it can mean significant energy consumption, which impacts battery life. Thus, in some example embodiments, the evaluation process can be performed on the gNB side, where the UE periodically transmits a large set of codewords, e.g., each codeword individually. The process of evaluating the decoder can require channel links with various conditions in order to perform different evaluations over different SNR ranges.

[0175] To enable the evaluation process, the UE can transmit and / or receive an evaluation request indicator (EQI) to / from the gNB to indicate that the UE is ready to communicate all-zero codewords in order to evaluate a particular NBP model. The EQI can be associated with an index of the NBP architecture and / or variant, and provide a padding pattern (ideally, an all-zero sequence) to help the gNB perform accurate estimation of the CQI. Further, the EQI can include or indicate parameters for the simulation, including a minimum number of errors that should be collected per SNR value and / or a minimum number of blocks that should be collected per SNR. The UE can then receive an evaluation ready indicator (ERI) from the gNB, which indicates that the gNB is ready to start receiving the evaluation codewords. The evaluation codewords can be modulated using BPSK / QPSK to keep the performance table small. However, in another example, higher order modulation and larger performance tables can be used if the capabilities allow. The UE can then start communicating all-zero codewords that are subsequently decoded by the gNB, while storing the number of errors and the index of the communicated codeword. When the stopping criteria for the evaluation is satisfied for a particular SNR, the UE can receive an evaluation completion indicator (ECI) indicating that the evaluation is complete for the given SNR value.

[0176] Note that the various embodiments above and discussed elsewhere herein provide several advantages and / or technical improvements. For example, the various embodiments can provide or facilitate a data-centric channel coding that depends on the learning environment, which results in higher transmission reliability (i.e., better performance) and lower decoding latency (i.e., requires fewer decoding iterations). Further, according to the various embodiments, online performance monitoring provides scalability of the NBP decoding AI / ML model, as well as adaptability to new channel and transmission scenarios and environments.

[0177] Some example embodiments can include methods for distributed training and / or evaluation of NBP decoders. For example, various embodiments can include methods for distributed UE training, where more than one UE (e.g., multiple or many UEs) participate in training a particular NBP decoding architecture / variant in a cooperative manner. Distributed training allows for a reduction in training computational complexity at the UE side by receiving fewer data sets, while providing diversity in SNR range by using different channel links through the same gNB. Additionally, NBP model training speed can be significantly increased as it is processed in parallel by the UEs.

[0178] In this example of distributed training, the gNB can initiate the training request TI by transmitting the training request TI to several UEs. For example, the gNB can transmit a periodic TI, where the UE can periodically receive the TI signaling from the gNB, the gNB can transmit an aperiodic TI, where the UE can receive the TI in an aperiodic manner, and / or the gNB can trigger a TI, where the UE can initiate the training by transmitting the TI to the gNB.

[0179] In the periodic and aperiodic cases, the UE can inform the gNB that it is available for training by transmitting a TCR, followed by a response from the gNB, for example, containing a training configuration, for example, using a training configuration signal (TCS). In a triggered TI, the UE can receive the TCR from the gNB and then can transmit the TCS, for example, containing the NBP and AI / ML parameters.

[0180] Generally, distributed training is asynchronous, where each UE independently trains its own receive data set. UEs that wish to participate in NBP training can transmit a TCR and then receive a TCS, which means that the UEs share the same NBP model initial weights and parameters. In some examples, the TCR can contain or indicate information related to the identity of the UE that transmitted the TCR.

[0181] In cooperative training, each participating UE can send the computed gradient to the gNB after each round. In this example, the gNB, rather than the UE, can be responsible for the gradient update and cost function update. Thus, the UE can be configured to send signaling indicating the current round and the computed gradient, which allows the gNB to adequately oversee the solution and cost function update.

[0182] Similar to the UE and joint training, after the training is complete, each participating UE can inform the gNB that the training is complete, for example, by sending a TCI signal.

[0183] The NBP evaluation at the gNB for a single UE can lack diversity, resulting in a long delay to update the model, as the gNB has to wait for other EQI requests to diversify the channel link when the UE emits different channel conditions. To overcome this (or similar) problem at least, the NBP evaluation can be collaborative, with different UEs participating with different channel links to the gNB. In this example, the EQI signal can be broadcast to different UEs wishing to participate in the decoding evaluation, as opposed to the above discussion. In this embodiment, the UEs can receive the EQI signal along with the NBP decoder index and Monte Carlo hyperparameters. Note that the new weights should not be shared, as the UEs only participate in the encoding process. Each UE can then send an ERI indicator to initiate the evaluation process. The process can then proceed as described above similar to the single UE evaluation process. When the evaluation is complete, each participating UE can expect an ECI signal from the gNB.

[0184] After the evaluation phase, to update the model and performance table, the UE can use a BLER reference signal (BRS) to receive a new BLER value along with the NBP model index, which helps the UE decide whether to update the NBP model weights or keep the old values. Since most of the evaluation is performed by the gNB, the gNB can report the results to the UE, and the UE can decide whether to update the trained weights. The update process can be based on the performance table available to the UE for different NBP models. For each NBP model, the performance table maps or correlates the BLER performance to SNR values. The UE can use the BRS to perform a comparison of the new BLER value received from the gNB and can then update the NBP ML model. After updating the NBP model and weights, the UE can report this update to the gNB, for example, by sending an update model indicator (UMI) signal via PUCCH to confirm that the model is updated by the newly trained weights.

[0185] Some example embodiments can include methods for SNR adaptive training using a parameter adapter neural network. In some examples, training the NBP can cause the ML model to overfit or underfit. For example, overfitting can result from overtraining the NBP at high SNR intervals, which enables the model to learn weights with low variance that are particularly well suited for that interval. Thus, the NBP model can be overfit and, at the inference phase, it can not perform optimally at lower SNR values.

[0186] On the other hand, training the NBP at low SNR intervals can hinder the model from effectively learning the weights, as the decoder typically cannot make correct decisions at very low SNR values. Thus, in this case, the model does not fit well enough and can not perform well on other SNR regimes.

[0187] Typically, training the NBP model in the waterfall region is a method to prevent overfitting and underfitting. However, the waterfall region can be different for each code and code rate / code length. The waterfall region is the interval of SNR where the BLER curve starts to drop. Finding the waterfall region for a particular coding configuration is not an easy task and can not be accurate. Also, the optimal NBP weights can be different for each particular SNR value and since the set of SNR values is infinite, it is generally not practical to store the weights for different SNR values.

[0188] The parameter adapter neural network (PAN) provides a good solution for enabling the NBP model to be independent of SNR by allowing the NBP to be trained on different SNR values while simultaneously retraining it, while training a separate shallow neural network (with one hidden layer) to adaptively scale the trained weights of the NBP model with respect to the SNR value. This SNR adaptive PAN is a neural network with one hidden layer, with one input unit, which is the SNR value, and a few hidden neurons with ReLU activation function, such that ReLU(x) = max(x, 0), and for the SS architecture, for each parameter output 1, which is most convenient for this configuration. The NBP-SS architecture can have only 2 weights, in addition to an optional weight that can be used to scale the LLR input. Thus, the UE can use different PANs to scale different parameters. For the relaxation factor γ, the activation function in the output layer can be a sigmoid function, since γ ∈ [0, 1]. For the weights associated with all edges, in addition to the scaling parameter for increasing the output scaling, the activation function can also be sigmoid, or the output layer can not have an activation function.

[0189] FIG. 6 Examples of PAN architectures that can be used for SNR adaptive NBP decoding are shown in accordance with various embodiments. In some embodiments, to train the PAN, the training process can be performed in accordance with two different options. In one option, the PAN can be trained together with the NBP model. In another option, the PAN can be trained after learning the NBP model weights (e.g., directly after learning the NBP model weights).

[0190] According to various embodiments, the UE can initiate PAN training by, for example, sending a parameter adapter training request (PATR) to the gNB via PUCCH, which indicates that the UE is ready to collect a certain data set at a certain SNR. In one embodiment, the PATR can be included in the TCR, where the UE indicates that the PAN should also be trained. For example, if the PATR is 0 (e.g., an indicator in the PATR is set to 0), then no PAN training is needed; otherwise, if the PATR is 1 (e.g., an indicator in the PATR is set to 1), then training the PAN is mandatory. The UE can then receive a PAN confirmation signal (PCS) from the gNB along with CSI, from which the UE is able to estimate the SNR value at which the PAN will be trained. The UE then starts receiving QPSK modulated all-zero codewords to train the PAN. In one embodiment, the PATR can be sent jointly with the TCR for joint PAN / NBP training. In another implementation, the PATR can be sent separately to indicate that the training is separate. Subsequent steps can be similar to the process of training the NBP AI / ML model, for example, as discussed above with respect to FIG. 4 and 5

[0191] According to certain embodiments, the methods described herein for SNR adaptive training of a PAN provide several technical advantages and / or improvements. For example, separate training of the PAN allows for the use of very low complexity and / or latency NBP models (NBP-SS-PAN) while outperforming traditional decoders. Additionally, the decoder has the ability to adapt its parameters to different channel conditions that are not necessarily observed during data set collection.

[0192] Some example embodiments can include methods for inference and user-centric selection of NBP models. The inference phase of the NBP decoder does not require implementing any DNN or AI / ML model. This is essentially due to the natural structure of the BP decoder, which is considered as message passing, while the basic BP becomes weighted BP, and the weights are those trained during the training phase.

[0193] The trained weights are different for each NBP architecture and variant. Therefore, in the inference phase, the UE can use different selection metrics to adequately select the NBP architecture for different use cases and scenarios.

[0194] The UE can store a lookup table that stores the trained weights associated with each NBP variant / architecture. Similarly, in one embodiment, the gNB can also store the same updated lookup table. Alternatively or additionally, in some embodiments, the weights can be sent from the UE to the gNB prior to the inference phase.

[0195] ​The selection metric procedure used by the UE to select the NBP for use can be based on one or more of the following criteria: latency, computational complexity, reliability, channel conditions, and / or code length / code rate.

[0196] The NBP architectures and variants provide different latencies. The decoding latency in the NBP grows with the number of edges of the training, the number of decoding iterations required, and the computational complexity required for the check node updates. In one embodiment, the UE can use a convenient NBP model based on its needs and available resources, including whether it tolerates decoding latency under the use case being considered.

[0197] In one embodiment, the UE can select an appropriate NBP model based on computational complexity, which can be associated with the battery life of the UE. Computational complexity is also related to the NBP architecture, where FF requires more complexity than RNN and its subsequent SS. Furthermore, this variant plays a key role as the simplified variants (NMS, NOMS) aim to reduce complexity. The UE can be able to dynamically select the NBP model based on its preferences and resources.

[0198] In one embodiment, the UE can be able to select the NBP model based on its reliability (i.e., BLER performance with respect to SNR intervals and channel models). The NBP models differ in performance and sometimes confuse performance in some specific SNR intervals. The UE can consider whether it is working in an error floor region, waterfall region, or error flat region according to the code structure and parameters. In this case, after estimating the channel conditions, the UE can decide which NBP model to adopt in order to improve transmission reliability.

[0199] The channel conditions criteria can also be related to reliability. Depending on the estimated SNR value, the UE can decide which NBP model to consider based on the code structure, code length, and / or code rate. The BLER curve behavior depends on the code, code length, code rate, and SNR. In one embodiment, based on the observed SNR, the UE can select the optimal NBP model that minimizes the BLER performance.

[0200] In one embodiment, the UE can select the NBP model based on the code length and code rate. The code length and code rate affect the size of the PCM used for decoding, and thus the number of trained weights associated with different edges. Therefore, this criterion can also be related to the latency, computational complexity, and reliability metrics, as the dimension of the PCM affects the complexity, latency, and performance. The UE can consider a combination of all of the above criteria in order to find the best trade-off.

[0201] It should be noted that for the reliability and channel conditions criteria, the UE’s decision can be based on a performance table associated with each NBP model. The UE can also select the NBP model by combining a set of the above selection metrics.

[0202] It can not be easy to illustrate the differences between NBP variants / architectures with respect to those metrics, and it can also be data driven, especially for reliability. The priority for optimizing reliability can be based on a stored performance table that contains a set of SNR values and their BLER performance for each NBP decoder. Typically, the FF architecture performs better in the low SNR region, and the SPA variant is also better than the others in low SNR. However, this behavior changes in the high SNR region, where the RNN architecture provides the same performance and sometimes better performance if properly trained. This can be the same for the NMS variant compared to SPA, which can sometimes outperform SPA at the cost of lower latency and complexity. For the SS-PAN architecture, the complexity and latency are lower at the cost of worse performance at low SNR, and a slight drop in performance in the high SNR region. For the code length metric, especially for LDPC codes, long length provides very satisfactory performance compared to short length. In fact, for long block length mechanisms, especially for cases and use cases where latency plays a key role in meeting the requirements, lower complexity / latency NBP variants / architectures can be a good choice.

[0203] In various embodiments, for the inference phase with SS-PAN architecture, the UE can receive the transmitted signal by the gNB, proceed by feeding the SNR of the estimated channel link to the PAN, then adjust its NBP weights according to the output of the PAN. Then, the UE can proceed by decoding the received sequence using the NBP decoder.

[0204] Therefore, the above exemplary embodiments advantageously allow the UE and the network to employ a user-centric channel coding approach, taking into account many metrics, and selecting the appropriate NBP AI / ML model that can meet the requirements of the UE or the network.

[0205] In view of the foregoing, various embodiments can be directed to a method implemented in a wireless transmit / receive unit (WTRU). The method can include receiving training configuration information from a network element, the training configuration information indicating: (i) an index of a neural belief propagation (NBP) variant or architecture and / or (ii) an update to one or more received hyperparameters for the NBP variant or architecture. The method can include sending first information to the network element indicating that the WTRU is ready to train the NBP variant or architecture, and receiving a training dataset from the network element including an all-zero codeword associated with a channel coding technique. The all-zero codeword can have a defined information block length and code rate. The method can then include starting training the NBP variant or architecture using the received training dataset and any of the hyperparameters. After one or more training iterations, the method can include sending second information to the network element indicating a cost function value associated with the trained NBP variant or architecture, and sending third information to the network element indicating trained weights associated with the trained NBP variant or architecture, if the training is completed.

[0206] One embodiment can relate to a WTRU including circuitry (e.g., a processor, memory, and / or a transceiver) configured to receive training configuration information from a network element, the training configuration information indicating: (i) an index of a neural belief propagation (NBP) variant or architecture and / or (ii) an update to one or more received hyperparameters for the NBP variant or architecture. The WTRU can be configured to send first information to the network element indicating that the WTRU is ready to train the NBP variant or architecture, and receive a training dataset from the network element including an all-zero codeword associated with a channel coding technique. The all-zero codeword can have a defined information block length and code rate. The WTRU can be configured to start training the NBP variant or architecture using the received training dataset and any of the hyperparameters. After one or more training iterations, the WTRU can be configured to send second information to the network element indicating a cost function value associated with the trained NBP variant or architecture, and send third information to the network element indicating trained weights associated with the trained NBP variant or architecture, if the training is completed.

[0207] In various embodiments, the method can include or the WTRU can be configured to receive configuration information indicating a lookup table associating NBP models and parameters. In other words, for example, a lookup table can associate NBP variants or architectures with hyperparameters.

[0208] In various embodiments, the NBP model can include a FNSPA, a RNSPA, a FNNMS, a RNNMS, an NBP-SS, and / or an NBP-SS-PAN.

[0209] In various embodiments, the method can include or the WTRU can be configured to receive a training indicator from a network element, where the training indicator includes any of an indication of a number of OFDM symbols, slots, and subframes before the training is to begin.

[0210] In various embodiments, the method can include or the WTRU can be configured to transmit a request to a network element to receive a training configuration. In some embodiments, the second information can be transmitted after each training iteration, after a predefined set of iterations, and / or when the training is complete.

[0211] In various embodiments, the training is completed (e.g., deemed complete) after all training iterations are completed or on a condition that a loss function of the NBP variant or architecture exceeds a predefined threshold.

[0212] In various embodiments, the method can include or the WTRU can be configured to receive a BLER reference signal (BRS) from a network element, where the BRS can indicate a pair of SNR / BLER values and / or an index of the NBP. In one embodiment, the method can include (or the WTRU is configured to) compare the SNR / BLER values to those in a stored BLER table, determine whether the weights of the NBP model should be updated based on the comparison, and transmit an indication of the update to the model to the network element if the weights are updated.

[0213] In various embodiments, the training can be performed in cooperation with at least one other WTRU (e.g., a plurality of WTRUs), or in cooperation between at least one other WTRU, and the third information can indicate the computed gradient and an index associated with the training iteration.

[0214] In various embodiments, the method can include or the WTRU can be configured to, when a stop criterion is met, receive or transmit an evaluation request indicator indicating any of an NBP model index and a Monte Carlo parameter associated with the NBP model; receive or transmit an evaluation ready indicator (ERI); transmit a set of codewords and their indices; and receive an evaluation completion indicator.

[0215] In various embodiments, the method can include or the WTRU can be configured to transmit a request to a network element to train a parameter adapter network (PAN), and receive a PAN configuration signal from the network element indicating any of a PAN architecture and parameters, and / or a CSI-RS for estimating a channel.

[0216] In various embodiments, the request to train the PAN can be sent with the request to receive the training configuration.

[0217] In various embodiments, the method can include or the WTRU can be configured to select the NBP variant or architecture using selection criteria based on any one or more of: latency, complexity, reliability, channel conditions, and coding parameters.

[0218] Although features and elements are provided in particular combinations in the above examples, one of ordinary skill in the art will appreciate that each feature or element can be used alone or in any combination with the other features and elements. In this disclosure, unless expressly provided otherwise, the description of an example should not be interpreted as an implication that elements of that example are required in every instance, and that every instance requires the same combination of elements. The disclosure is not to be limited in scope by the specific examples described herein, which are intended for the purpose of exemplification. Many modifications and variations will occur to those of ordinary skill in the art upon reading of this disclosure, which are intended to be within the scope of the claims. The description and drawings are to be regarded as illustrative in nature and not as restrictive. The disclosure is not limited to the specific methods or systems described herein, but only to the full scope of equivalents of the claims.

[0219] In some example embodiments described herein, the (e.g., configuration) information can be described as being received by the WTRU from the network, e.g., through system information or via any type of protocol message. Although not explicitly mentioned in the embodiments described herein, the same (e.g., configuration) information can be pre-configured in the WTRU (e.g., via any kind of pre-configuration method, e.g., via factory settings), such that the (e.g., configuration) information can be used by the WTRU without the need to be received from the network.

[0220] Any of the features, variations, or embodiments described for the method are compatible with devices comprising means for processing the disclosed method, such as devices comprising a processor configured to process the disclosed method, computer program products comprising program code instructions, and non-transitory computer-readable storage media storing program instructions.

[0221] For simplicity, the foregoing embodiments are discussed in terms of terminology and structures for devices having infrared capabilities, i.e., infrared emitters and receivers. However, the embodiments discussed are not limited to these systems, but can be applied to other systems using other forms of electromagnetic waves or non-electromagnetic waves such as acoustic waves.

[0222] It is also to be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting. As used herein, the term "video" or the term "image" can mean any one of a snapshot, a single image, and / or a plurality of images displayed on a time basis. As another example, the term "user equipment" and its acronym "UE," the term "remote," and / or the term "head-mounted display" or its acronym "HMD," when referenced herein, can mean or include (i) a wireless transmit and / or receive unit (WTRU); (ii) any one of a plurality of embodiments of a WTRU; (iii) a device with wireless capability and / or wired capability (e.g., tetherable) that is configured with some or all of the structure and functionality of a WTRU, among others; (iii) a device with wireless capability and / or wired capability that is configured with less than all of the structure and functionality of a WTRU; or (iv) the like. Reference is made herein to FIGS. 1A-1D Details of example WTRUs are provided, which can represent any of the WTRUs set forth herein. As another example, various embodiments disclosed above and below herein are described as utilizing a head-mounted display. Those skilled in the art will recognize that devices other than head-mounted displays can be utilized, and some or all of the various disclosed embodiments can be modified accordingly without undue experimentation. Examples of such other devices can include drones or other devices configured to stream information for providing an adapted reality experience.

[0223] In addition, the methods provided herein can be implemented in a computer program, software, or firmware incorporated in a computer- readable medium for execution by a computer or processor. Examples of computer-readable media include electronic signals (optical, electrical or the like) and computer- readable storage media. Examples of computer-readable storage media include, but are not limited to, removable floppy disks, RAM disks, solid state RAM, magnetic tape, hard disk drives, optical storage, and the like. The processor associated with software can be for execution of the software by a WTRU, UE, terminal, base station, RNC, or any host computer. The software associated with the processor can be for implementing the radio frequency transceiver in the WTRU, UE, terminal, base station, RNC, or any host computer.

[0224] Variations of the methods, apparatus and systems provided above are possible without departing from the scope of the present disclosure. Given the wide variety of embodiments that can be implemented, it should be understood that the illustrated embodiments are only examples and should not be taken as limiting the scope of the claims. For example, the embodiments provided herein include a handheld device that can include or use with any suitable voltage source (such as a battery or the like) that provides any suitable voltage.

[0225] Further, in the embodiments provided above, processing platforms, computing systems, controllers, and other devices including processors are noted. These devices can include at least one central processing unit ("CPU") and memory. In accordance with the practices of persons skilled in the art of computer programming, reference to acts and symbolic representations of operations or instructions can be performed by the various CPUs and memories. Such acts and operations or instructions can be referred to as being "executed," "computer executed" or "CPU executed"

[0226] Those skilled in the art and informed by the disclosure will understand that the acts and symbolic representations of operations or instructions include the manipulation of electrical signals by the CPU. The electrical system representations, such as a graphical representation of data bits, can result in transformations or reductions of the electrical signals or maintenance of the data bits at memory locations in the memory systems by the storage of different bits at a different memory location resulting in a shift in the nature and embodiment of the signals, which represents all of the steps of some exemplary embodiments described herein. The maintenance of data bits at memory locations includes a manipulation of electrical signals analog or digital to maintain and transform the memory locations in the CPU and its operations, thereby changing the operation of the CPU and processing of the signals.

[0227] The data bits can also be maintained on computer-readable media including magnetic disks, optical disks, and any other volatile (e.g., Random Access Memory (RAM)) or non-volatile (e.g., Read Only Memory (ROM)) mass storage system readable by the CPU. The computer-readable medium can include cooperating or interconnected computer-readable media, which exist exclusively on the processing system, or are distributed among multiple interconnected processing systems that can be local or remote to the processing system. It is understood that the embodiments are not limited to the above-mentioned memory, and that other platforms and memory can support the provided methods.

[0228] In illustrative embodiments, any of the operations, processes, etc. described herein can be implemented as computer-readable instructions stored on a computer-readable medium. The computer-readable instructions can be executed by a processor of a mobile unit, network element, and / or any other computing device.

[0229] There is little distinction between the use of hardware and software in the various aspects of systems. The use of hardware and software is generally (but not always, in that in certain contexts the choice between hardware and software can become significant) a design choice representing cost vs. efficiency tradeoffs. There can be various vehicles by which processes and / or systems and / or other technologies described herein can be effected (e.g., hardware, software, and / or firmware), and the preferred vehicle can vary with the context in which the processes and / or systems and / or other technologies are deployed. For example, if an implementer determines that speed and accuracy are paramount, the implementer can opt for mostly hardware and / or firmware implementations. If flexibility is paramount, the implementer can opt for software implementations. Alternatively, the implementer can opt for some combination of hardware, software, and / or firmware.

[0230] The foregoing detailed description has set forth various embodiments of the devices and / or processes via the use of block diagrams, flowcharts, and / or examples. Insofar as such block diagrams, flowcharts, and / or examples contain one or more functions and / or operations, it will be understood by those within the art that each function and / or operation within such block diagrams, flowcharts, and / or examples can be implemented, individually and / or collectively, by a wide range of hardware, software, firmware, or virtually any combination thereof. In one embodiment, several portions of the subject matter described herein can be implemented via Application Specific Integrated Circuits (ASICs), Field Programmable Gate Arrays (FPGAs), digital signal processors (DSPs), and / or other integrated formats. However, those skilled in the art will recognize that some aspects of the embodiments disclosed herein, in whole or in part, can be equivalently implemented by circuits, combined circuits, and / or other functional paradigms; the preferred implementations are thus, in terms of the functional operations disclosed, not limited to the method steps presented in this description and / or the sequence of blocks presented and / or equivalent hardware structures. Skilled artisans will appreciate that these functional steps can typically be performed in the sequence shown, but variations of this sequence can also be performed and / or other functional steps can be added or performed. Those skilled in the art will appreciate that the order of execution or the specific functions of any process embodiment can be varied for each particular implementation. Accordingly, other embodiments are within the scope of the following claims, and outside the scope of these claims, the scope being determined by the body of the claims and the equivalence thereto.

[0231] Those skilled in the art will recognize that, in the context of this specification, devices and / or processes are often described in terms of sequences of actions to be performed by, for example, elements of a computing device. It will be recognized that various elements of the devices and / or processes could be performed by specialized computers or processors as described above. It will also be recognized that such devices and / or processes might be embodied in whole or in part by circuitry, by computer hardware, firmware, software, or combinations thereof. To clearly illustrate this interchangeability of hardware and software, various components will be described herein as being stored in and / or executed by various elements of a computing device. But it will be understood that such components can be implemented by specialized computers or processors that are specifically programmed, configured, and / or constructed to perform the tasks described herein. Moreover, it will be understood that various software algorithms or computer programs that embody the processes described herein can be stored in memory and / or implemented in hardware, including but not limited to various application-specific processors. In fact, computer programs that embody the processes described herein are typically stored in a non-transitory computer-readable storage medium at some point during the computer life cycle. Since such computer programs are a form of machine-readable instructions, they can be read onto a machine's volatile memory (e.g., random access memory within a computer) to alter the machine's operation, e.g., by configuring the processor(s) to execute processes described herein. After the machine operates in this altered manner, the computer program in volatile memory can be discarded, or it can be stored in non-volatile memory (e.g., a hard drive, or flash memory) either before or after discarding. Accordingly, a machine-readable medium can include any non-transitory medium that is capable of storing or encoding a sequence of instructions for execution by the machine and that causes the machine to perform any one of the methodologies of the present disclosure. The machine-readable medium can include, but is not limited to, solid-state memories, optical media, magnetic media, and / or signals.

[0232] The herein described subject matter sometimes illustrates different components contained within, or connected with, different other components. It is to be understood that such depicted architectures are merely examples, and that in fact many other architectures can be implemented which achieve the same functionality. In a conceptual sense, any arrangement of components to achieve the same functionality is effectively "associated" such that the desired functionality is achieved. Hence, any two components herein combined to achieve a particular functionality can be seen as "associated with" each other such that the desired functionality is achieved, irrespective of architectures or intermediate components. Likewise, any two components so associated can also be viewed as being "operably connected", or "operably coupled", to each other to achieve the desired functionality, and any two components capable of being so associated can also be viewed as being "operably couplable", to each other to achieve the desired functionality. Specific examples of operably couplable include but are not limited to physically mateable and / or physically interacting components and / or wirelessly interactable and / or wirelessly interacting components and / or logically interacting and / or logically interactable components.

[0233] With respect to the use of substantially any plural and / or singular terms herein, those having skill in the art can translate the plural and / or singular terms to a singular and / or plural form, as appropriate. For the sake of clarity, various singular / plural permutations of the terms can be expressly set forth herein.

[0234] Those skilled in the art will understand that, in general, the terms used herein, and particularly in the appended claims (e.g., the bodies of the appended claims), are generally intended as "open-ended" terms (e.g., the term "including" should be interpreted as "including but not limited to," the term "having" should be interpreted as "having at least," the term "comprising" should be interpreted as "including, but not limited to," etc.). Those skilled in the art will also understand that if a specific number of an introduced claim recitation is intended, such intent will be explicitly recited in the claim, and in the absence of such recitation, no such intent is present. For example, where only one item is intended, the term "single" or similar language may be used. To aid understanding, the following appended claims and / or the description herein may include the use of the introductory phrases "at least one" and "one or more" to introduce claim recitations. However, the use of such phrases should not be interpreted as implying that a claim recitation introduced by the indefinite article "a" or "an" limits any particular claim including such introduced claim recitation to embodiments including only one such recitation, even when the same claim includes the introductory phrases "one or more" or "at least one" and an indefinite article such as "a" or "an" (e.g., "a" and / or "an" should be interpreted as meaning "at least one" or "one or more"). The same applies to the use of definite articles to introduce claim recitations. In addition, even if a specific number of an introduced claim recitation is explicitly recited, those skilled in the art will recognize that such recitation should be interpreted to mean at least the recited number (e.g., the unmodified recitation of "two recitations" means at least two recitations, or two or more recitations, in the absence of other modifiers). Furthermore, in those instances where a convention similar to “at least one of A, B, and C, etc.” is used, such construction is generally intended to be in the sense that one skilled in the art would understand the convention (e.g., “a system having at least one of A, B, and C” would include but is not limited to systems having only A, only B, only C, A and B, A and C, B and C, and / or A, B, and C, etc.). In those instances where a convention similar to “at least one of A, B, or C, etc.” is used, such construction is generally intended to be in the sense that one skilled in the art would understand the convention (e.g., “a system having at least one of A, B, or C” would include but is not limited to systems having only A, only B, only C, A and B, A and C, B and C, and / or A, B, and C, etc.). One skilled in the art would also understand that any disjunctive conjunctions and / or phrases, whether in the specification, claims, or drawings, that actually present two or more alternative terms should be understood to contemplate the possibility of including one, either, or both of these terms.Further, as used herein, the term "any of" followed by a listing of a plurality of items and / or categories of items is intended to mean "any of the item s and / or categories of items" but not excluding combinations of the items and / or categories of items. Further, as used herein, the term "set" is intended to mean any number of items, including zero. Additionally, as used herein, the term "number" is intended to mean any number of items, including zero. And as used herein, the term "plurality" is intended to mean "more than one."

[0235] Further, where features or aspects of the disclosure are described in terms of Markush groups, a person of ordinary skill in the art will recognize that the disclosure is also thereby described in terms of any individual member or subgroup of members of the Markush group.

[0236] As will be understood by those of ordinary skill in the art, all ranges disclosed herein are also intended to encompass any and all possible sub-ranges and combinations of sub-ranges thereof, for any and all purposes. Any listed range can be easily reduced to, and is intended to include, a half, third, quarter and fifth of the upper value of the range, and every fraction thereof. As a non-limiting example, each range discussed herein can be easily reduced to a lower third, middle third and upper third, and the like. A person of ordinary skill in the art will also understand that all language such as "up to," "at least," "greater than," "less than," and the like include the integer that the term is used with. Finally, a range includes each individual member. Thus, for example, a group of 1 to 3 units refers to a group of 1, 2, or 3 units. Similarly, a group of 1 to 5 units refers to a group of 1, 2, 3, 4, or 5 units, and so forth.

[0237] Further, no limitation from the order or elements provided should be inferred, unless specifically stated otherwise. Additionally, use of the term "means" for any claim in accordance with 35 U.S.C. § 112, 6 or means-plus-function claim format requires support from a claim reciting functions followed by structure, means, or a combination of functions and structures.

[0238] While various embodiments have been described in terms of a communication system, it is contemplated that these systems can be implemented in software on a microprocessor / general purpose computer (not shown). In certain embodiments, one or more of the functions of the various components can be implemented in software that controls a general purpose computer.

[0239] In addition, although some example embodiments are shown and described herein, the present application is not limited to the details of the examples. Instead, various modifications and changes can be made within the scope and equivalents of the claims, without departing from the spirit or scope of the application.

[0240] References

[0241] The above can have referenced the following references, each of which is incorporated herein by reference in its entirety:

[0242] 3GPP TS 38.212, Multiplexing and channel coding, v16.0.0;

[0243] Nachmani, Eliya, et al. “Learning to decode linear codes using deep learning.” 2016 54th Annual Allerton Conference on Communication, Control, and Computing (Allerton): 341-346;

[0244] E. Nachmani, E. Marciano, L. Lugosch, W. J. Gross, D. Burshtein, and Y. Be’ery, “Deep learning methods for improved decoding of linear codes,” in IEEE Journal of Selected Topics in Signal Processing, vol. 12, no. 1, pp. 119-131, Feb 2018, doi: 10.1109 / JSSTS.2017.2788405.

[0245] L. Lugosch and W. J. Gross, “Learning from adjoints,” in Asilomar Conference on Signals, Systems, and Computers, “Machine Learning for Wireless Systems” Invited Session, Oct 2018.

[0246] L. Lugosch and W. J. Gross, “Neural shift-min sum decoding,” in IEEE International Symposium on Information Theory (ISIT), Aachen, Germany, 2017, pp. 1361-1365; and

[0247] Dai, J., Tan, K., Si, Z., Niu, K., Chen, M., Vincent Poor, H., and Cui, S. (2021). Learning decoding methods for protograph LDPC codes. IEEE Journal of Selected Topics in Communications, 39(7), 1983-1999. [9427170].

[0248] Partial glossary of abbreviations and acronyms

[0249] ACK acknowledgement

[0250] ADAM adaptive moment estimation

[0251] AE autoencoder

[0252] AI artificial intelligence

[0253] AIML artificial intelligence / machine learning

[0254] BCH Bose-Chaudhuri-Hocquenghem

[0255] BG base graph

[0256] BLER block error rate

[0257] BP-MS belief propagation min-sum

[0258] BP-SPA belief propagation sum-product algorithm

[0259] CA-SCL CRC-aided serial cancellation list

[0260] CN check node

[0261] CRC cyclic redundancy check

[0262] CSI channel state information

[0263] D2D device-to-device (e.g., LTE sidelink)

[0264] DL downlink

[0265] eMBB enhanced mobile broadband

[0266] FEC forward error correction

[0267] FF feedforward

[0268] FNSPA feedforward neural sum-product algorithm

[0269] gNB NR NodeB

[0270] HARQ hybrid automatic repeat request

[0271] LDPC low-density parity check

[0272] LLR log-likelihood ratio

[0273] MCS modulation and coding scheme

[0274] MIMO multiple-input multiple-output

[0275] ML machine learning

[0276] MTC machine type communication

[0277] NACK negative acknowledgement

[0278] NBP neural belief propagation

[0279] NNMS neural network normalized minimum sum

[0280] NOMS neural offset minimum sum

[0281] NR new radio

[0282] OSD ordered statistics decoding

[0283] PCM parity check matrix

[0284] PHY physical layer

[0285] PRB physical resource block

[0286] PR-LDPC protograph low-density parity-check

[0287] ReLU rectified linear unit

[0288] RMSPROP root mean square propagation

[0289] RM Reed-Muller

[0290] RNN recurrent neural network

[0291] RNSPA recurrent neural and sum-product algorithm

[0292] RS reference signal

[0293] SGD stochastic gradient descent

[0294] TRx transceiver

[0295] UL uplink

[0296] URLLC ultra-reliable low-latency communication

[0297] VN variable node

Claims

1. A method implemented in a wireless transmit / receive unit (WTRU), the method comprising: receiving training configuration information from a network element, the training configuration information indicating any of: (i) an index of a neural belief propagation (NBP) variant or architecture, and (ii) an update to one or more received hyperparameters for the NBP variant or architecture; sending first information to the network element, the first information indicating that the WTRU is ready to train the NBP variant or architecture; receiving a training data set comprising all-zero codewords associated with a channel coding technique, the all-zero codewords having a defined information block length and code rate; Starting training of the NBP variant or architecture using the received training dataset and any of the hyperparameters; After one or more training iterations, sending second information to the network element, the second information indicating a cost function value associated with the trained NBP variant or architecture; as well as In case the training is completed, third information is sent to the network element, the third information indicating trained weights associated with the trained NBP variant or architecture.

2. The method according to claim 1, comprising: Configuration information indicating a lookup table associating the NBP variant or architecture with hyperparameters is received.

3. The method according to any one of claims 1 to 2, wherein the NBP variant or architecture includes any one of the following: a feedforward neural sum-product algorithm, a recurrent neural sum-product algorithm, a feedforward neural network normalized minimum sum, a recurrent neural network normalized minimum sum, an NBP synchronization signal (SS), an NBP-SS parameter adapter network (PAN).

4. The method according to any one of claims 1 to 3, comprising: A training indicator is received from the network element, wherein the training indicator comprises any of: an indication of a number of orthogonal frequency division multiplexing (OFDM) symbols, time slots, and / or subframes before the training begins.

5. The method according to any one of claims 1 to 4, comprising: A request is transmitted to the network element to receive the training configuration.

6. The method according to any one of claims 1 to 5, wherein The second information is sent after any of: each training iteration, a predetermined set of iterations, or when training is completed.

7. The method according to any one of claims 1 to 6, wherein the training is completed after all the training iterations are completed or under the condition that the loss function of the NBP variant or architecture exceeds a predefined threshold.

8. The method according to any one of claims 1 to 7, comprising: receiving a block error rate (BLER) reference signal (BRS) from the network element, the BRS indicating any one of (i) a paired signal-to-noise ratio (SNR) / BLER value and (ii) an index of the NBP; comparing the SNR / BLER value with those in a stored BLER table; Based on the comparison, determining whether weights of the NBP variant or architecture should be updated; as well as In case the weights are updated, an indication of the updated weights is transmitted to the network element.

9. The method of any one of claims 1 to 8, wherein the training is performed in collaboration with or between at least one other WTRU, and wherein the third information indicates a calculated gradient and an index associated with a training iteration.

10. The method according to any one of claims 1 to 9, comprising: receiving or sending an evaluation request indicator indicating any one of an index associated with the NBP variant or architecture and a Monte Carlo parameter associated with the NBP variant or architecture; Receive or send an Evaluation Readiness Indicator (ERI); sending a set of codewords and indexes of the codewords; as well as When the stopping criteria are met, an evaluation completion indicator is received.

11. The method according to any one of claims 1 to 10, comprising: sending a request to the network element to train a parameter adapter network (PAN); as well as PAN configuration information is received from the network element, the PAN configuration information indicating any one of: a PAN architecture and parameters, and a channel state information reference signal (CSI-RS) for estimating the channel.

12. The method according to claim 11, wherein The request to train the PAN is sent together with the request to receive the training configuration.

13. The method according to any one of claims 1 to 12, comprising: The NBP variants or architectures were selected using selection criteria based on any of the following: Delay, complexity, reliability, channel conditions, and Encoding parameters.

14. A wireless transmit / receive unit (WTRU), the WTRU comprising: A circuit, the circuit comprising any one of a transceiver, a processor, and a memory, the circuit being configured to: receiving training configuration information from a network element, the training configuration information indicating any of: (i) an index of a neural belief propagation (NBP) variant or architecture, and (ii) an update to one or more received hyperparameters for the NBP variant or architecture; sending first information to the network element, the first information indicating that the WTRU is ready to train the NBP variant or architecture; receiving a training data set comprising all-zero codewords associated with a channel coding technique, the all-zero codewords having a defined information block length and code rate; Starting training of the NBP variant or architecture using the received training dataset and any of the hyperparameters; After one or more training iterations, sending second information to the network element, the second information indicating a cost function value associated with the trained NBP variant or architecture; as well as In case the training is completed, third information is sent to the network element, the third information indicating trained weights associated with the trained NBP variant or architecture.

15. The WTRU of claim 14 , wherein the circuit is configured to: Configuration information indicating a lookup table associating the NBP variant or architecture with hyperparameters is received.

16. The WTRU according to any one of claims 14 to 15, wherein the NBP variant or architecture includes any one of the following: a feedforward neural sum-product algorithm, a recurrent neural sum-product algorithm, a feedforward neural network normalized minimum sum, a recurrent neural network normalized minimum sum, an NBP synchronization signal (SS), an NBP-SS-parameter adapter network (PAN).

17. The WTRU of any one of claims 14 to 16, comprising: A training indicator is received from the network element, wherein the training indicator comprises any of: an indication of a number of orthogonal frequency division multiplexing (OFDM) symbols, time slots, and / or subframes before the training begins.

18. The WTRU of any one of claims 14 to 17, comprising: A request is transmitted to the network element to receive the training configuration.

19. The WTRU of any one of claims 14 to 18, wherein the second information is sent after any of: each training iteration, a predetermined set of iterations, or when training is completed.

20. The WTRU of any one of claims 14 to 19, wherein the training is completed after all the training iterations are completed or if a loss function of the NBP variant or architecture exceeds a predefined threshold.

21. The WTRU of any one of claims 14 to 20, comprising: receiving a block error rate (BLER) reference signal (BRS) from the network element, the BRS indicating any one of (i) a paired signal-to-noise ratio (SNR) / BLER value and (ii) an index of the NBP; comparing the SNR / BLER value with those in a stored BLER table; Based on the comparison, determining whether weights of the NBP variant or architecture should be updated; as well as In case the weights are updated, an indication of the updated weights is transmitted to the network element.

22. The WTRU according to any one of claims 14 to 21, wherein the training is performed in collaboration with at least one other WTRU or in collaboration between at least one other WTRU, and wherein the third information indicates a calculated gradient and an index associated with a training iteration.

23. The WTRU of any one of claims 14 to 22, comprising: receiving or sending an evaluation request indicator indicating any one of an index associated with the NBP variant or architecture and a Monte Carlo parameter associated with the NBP variant or architecture; Receive or send an Evaluation Readiness Indicator (ERI); sending a set of codewords and indexes of the codewords; as well as When the stopping criteria are met, an evaluation completion indicator is received.

24. The WTRU of any one of claims 14 to 23, comprising: sending a request to the network element to train a parameter adapter network (PAN); as well as PAN configuration information is received from the network element, the PAN configuration information indicating any one of: a PAN architecture and parameters, and a channel state information reference signal (CSI-RS) for estimating the channel.

25. The WTRU of claim 24, wherein the request to train the PAN is sent along with the request to receive the training configuration.

26. The WTRU of any one of claims 14 to 25, wherein the circuitry is configured to select the NBP variant or architecture using a selection criterion based on any one of: Delay, complexity, reliability, channel conditions, and Encoding parameters.

Images