Apparatus and method for distributed adaptive learning in wireless systems
By introducing adaptive learning agent entities into wireless communication systems, the computational complexity and output data size of machine learning models are dynamically adjusted, solving the performance degradation problem of distributed AI/ML schemes in dynamic wireless environments and improving network performance and resource utilization efficiency.
Patent Information
- Application Number
- CN202380100816.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2023-08-08
- Publication Date
- 2026-02-27
AI Technical Summary
Existing distributed AI/ML solutions struggle to adapt to changes in dynamic and unreliable wireless environments, leading to performance degradation and limited scalability.
A proxy entity is provided for adaptive learning machine learning models, which can dynamically adjust the computational complexity and output data size of the model according to computing and communication resources, and interact with the controller entity through a wireless communication channel to achieve adaptive learning.
It enables effective adjustment of the computational complexity and output data size of ML models in dynamic wireless environments, improves network performance, adapts to changes in wireless channels and computing power, reduces power consumption for resource-constrained devices, and optimizes communication resource allocation.
Smart Images
Figure CN121586901A_ABST
Abstract
Description
Technical Field
[0001] This disclosure relates to wireless communication. More specifically, this disclosure relates to devices and methods for distributed adaptive learning in wireless communication systems. Background Technology
[0002] Artificial intelligence (AI) and machine learning (ML) are being explored for use cases in 3GPP wireless communication systems that require collaboration between existing and new network nodes, such as between user equipment (UE) and base stations (BS), and between sensor-capable collaborative drones or mobile robots. For example, 3GPP is exploring the use of sensor-capable mobile robots as network nodes to support machine sensing and communication, with use cases including, but not limited to, factories, e-health, smart cities, and hazardous environments. Such network nodes can be powered by AI / ML and typically require a wireless link to a central node (controller) for coordination.
[0003] During training and perturbation, distributed AI / ML schemes such as Split Learning (SL) or Federated Learning (FL) typically operate in dynamic and unreliable wireless environments, with network nodes exhibiting time-varying states. To adapt distributed AI / ML schemes to dynamic wireless environments, methods have been proposed for storing and managing multiple ML models (each with different compression and complexity capabilities), requiring the system to select, load, and deploy appropriate ML models as the wireless environment changes. In highly dynamic wireless environments, this is neither practical nor scalable. Summary of the Invention
[0004] The purpose of this disclosure is to provide improved devices and methods for distributed adaptive learning in wireless communication systems.
[0005] The foregoing and other objectives are achieved through the subject matter of the technical solution in the independent claims. Other implementations are apparent from the dependent claims, the specification, and the drawings.
[0006] According to a first aspect, a proxy entity for adaptive learning is provided. The proxy entity is used to run a machine learning (ML) model for adaptive learning, wherein the ML model processes input data into output data with selectable, tunable computational complexity and selectable, tunable output data size. Furthermore, the proxy entity is used to estimate the current computational resources available for running the ML model and to obtain information indicating a selectable size for the output data of the ML model. The proxy entity is also used to select the computational complexity of the ML model and / or the size of the output data based on the estimate of the proxy entity's current computational resources and / or the information indicating a selectable size for the output data of the ML model.
[0007] Therefore, according to the first aspect, the agent entity can adapt its ML model to its computational capabilities (and possibly other communication resource capabilities) in a wireless communication system where collaboration between agent entities is required and the machine learning model is distributed among multiple agent entities. For example, according to the first aspect, the agent entity can adapt to limited and time-varying wireless resources and time-varying wireless channels between cooperating network nodes in a distributed ML model task. These changes can occur very quickly, for example, as quickly as the temporal coherence of the wireless channel. Furthermore, according to the first aspect, the agent entity can adapt to time-varying computational capabilities, for example, caused by contention between other different tasks running on the agent entity. For resource-constrained agent entities, such as mobile devices, UAVs, mobile robots, etc., running an ML model to the desired target accuracy may involve a large amount of computation, which can directly affect their power consumption. If the distributed ML model is not adaptively adapted, changes in the system may negatively impact network performance or the correct operation of the cooperating network nodes involved in the distributed ML model task. Moreover, wireless resources can be dynamically used in a shared channel among multiple agent entities, allowing more communication resources to be allocated to those agent entities experiencing degraded wireless channels. For example, this can be used for dynamic resource allocation of control channels that report channel state information.
[0008] In another possible implementation, the proxy entity is used to receive information from the controller entity via a wireless communication channel indicating the selectable size of the output data of the ML model. This allows the controller entity to centrally control the selectable size of the output data of the ML models of multiple proxy entities.
[0009] In another possible implementation, to obtain information indicating the selectable size of the output data of the ML model, the proxy entity estimates the current communication resources for communicating with the controller entity via a wireless communication channel. The proxy entity then selects the computational complexity of the ML model and / or the size of the output data based on the estimates of the current computational and communication resources. This allows the complexity of the ML model and / or the output data of the proxy entity to be adjusted according to the proxy entity's current computational and communication capabilities.
[0010] In another possible implementation, to estimate current communication resources, the agent entity determines the channel state information of the wireless communication channel between the agent entity and the controller entity, and selects the computational complexity of the ML model and / or the size of the output data based on the estimated current computational resources and the channel state information. This enables the agent entity to efficiently estimate its current communication capabilities.
[0011] In another possible implementation, the agent entity is also used to send the output data of the ML model to the controller entity via a wireless communication channel. This enables the controller entity to collect and process output data from multiple agent entities.
[0012] In another possible implementation, in response to sending the output data of the ML model to the controller entity, the agent entity is also configured to receive response data from the controller entity, wherein the response data is based on the output data of the agent entity's ML model and multiple other output data of multiple other ML models from multiple other agent entities. This enables the agent entity to receive feedback data from the controller entity based on output data from multiple agent entities.
[0013] In another possible implementation, the response data contains information instructing the agent entity to take action and / or information for performing backpropagation to update the agent entity's ML model. This enables the agent entity to perform actions and / or adjust its ML model based on feedback from the controller entity.
[0014] In another possible implementation, the agent entity is a user equipment used to exchange data with the controller entity via a wireless communication channel and a base station.
[0015] In another possible implementation, the ML model is the encoding part of an autoencoder, wherein the input data of the encoding part of the autoencoder is channel state information, and the output data of the encoding part of the autoencoder is compressed channel state information. This enables the agent entity to efficiently compress the channel state information based on the agent entity's current computing and / or communication resources.
[0016] In another possible implementation, the proxy entity is a mobile micro base station.
[0017] In another possible implementation, the agent entity is the base station, and the controller entity is the user equipment.
[0018] In another possible implementation, the ML model includes multiple processing layers for processing input data into output data, and wherein, in order to select the computational complexity of the ML model, a proxy entity is used to select a selectable number of processing layers from the multiple processing layers of the ML model. This allows the proxy entity to efficiently adjust the computational complexity of the proxy entity's ML model.
[0019] In another possible implementation, the proxy entity includes a battery for powering one or more processors of the proxy entity to implement the ML model, wherein the proxy entity estimates its current computing resources based on the battery's state of load. This allows the proxy entity to efficiently estimate its current computing resources for running the ML model.
[0020] According to the second aspect, a method for manipulating a proxy entity for adaptive learning is provided. The method includes the following steps:
[0021] Run a machine learning (ML) model for adaptive learning, wherein the ML model is used to process input data into output data with selectable, adjustable computational complexity and selectable, adjustable output data size;
[0022] Estimate the current computing resources of the agent entity used to run the ML model;
[0023] Obtain information indicating the selectable size of the output data of the ML model;
[0024] The computational complexity of the ML model and / or the size of the output data are selected based on an estimate of the agent entity’s current computing resources and / or information indicating the selectable size of the ML model’s output data.
[0025] The method according to the second aspect of this disclosure can be performed by a robot according to the first aspect of this disclosure. Therefore, other features of the method according to the second aspect of this disclosure are directly derived from the functionality of the robot according to the first aspect of this disclosure and its different implementations described above and below.
[0026] According to a third aspect, a computer program product is provided, including a computer-readable storage medium for storing program code, which, when executed by a computer or processor, causes the computer or processor to perform the method according to the second aspect.
[0027] The accompanying drawings and the following description illustrate details of one or more embodiments. Other features, objects, and advantages are clearly shown in the specification, drawings, and claims. Attached Figure Description
[0028] The embodiments of this disclosure will be described in more detail below with reference to the accompanying drawings, in which:
[0029] Figure 1 This is a schematic diagram illustrating multiple agent entities communicating with base station and controller entities for distributed adaptive learning according to an embodiment;
[0030] Figure 2 This is a table illustrating multiple ML model execution strategies defined for different conditions of the agent entity according to an embodiment;
[0031] Figure 3 This is a signaling diagram illustrating the dynamic adaptation of an ML model of a proxy entity for changing the conditions of the proxy entity according to an embodiment.
[0032] Figure 4 This is a signaling diagram illustrating the interaction between a base station controller entity and a UE agent entity for uplink transmission of compressed downlink channel state information according to an embodiment.
[0033] Figure 5 This is a signaling diagram illustrating the interaction between a base station agent entity and a UE controller entity for downlink transmission of compressed uplink channel state information according to an embodiment.
[0034] Figure 6 This is a signaling diagram illustrating the interaction between a controller entity and multiple micro base station agent entities according to an embodiment for coordinating multiple micro base station agent entities;
[0035] Figure 7 This is a flowchart illustrating a method for operating a proxy entity according to an embodiment of distributed adaptive learning.
[0036] In the following text, the same reference numerals refer to the same or at least functionally equivalent features. Detailed Implementation
[0037] In the following description, reference is made to the accompanying drawings, which form part of this disclosure, illustrating by way of description specific aspects of embodiments of this disclosure or aspects that may be used with respect to embodiments of this disclosure. It should be understood that embodiments of this disclosure may be used in other aspects and may include structural or logical variations not depicted in the drawings. Therefore, the following detailed description should not be construed in a limiting sense, and the scope of this disclosure is defined by the appended claims.
[0038] For example, it should be understood that the disclosure relating to the described method is also applicable to the corresponding device or system for performing the method, and vice versa. For instance, if one or more specific method steps are described, the corresponding device may include one or more units, such as functional units, to perform the described one or more method steps (e.g., one unit performs one or more steps, or multiple units perform one or more of a plurality of steps respectively), even if such one or more units are not explicitly described or shown in the drawings. Furthermore, if a particular apparatus is described according to one or more units (e.g., functional units), the corresponding method may include a step to perform the function of one or more units (e.g., one step performs the function of one or more units, or multiple steps perform the function of one or more of a plurality of units respectively), even if such one or more steps are not explicitly described or shown in the drawings. Moreover, it should be understood that, unless otherwise expressly stated, features of the various exemplary embodiments and / or aspects described herein can be combined with each other.
[0039] Before describing the detailed embodiments below, some terms will be introduced using one or more of the following abbreviations:
[0040] Machine Learning (ML)
[0041] AI (Artificial Intelligence)
[0042] MLOL (Machine Learning Orchestrator)
[0043] LLA Local Learning Agent
[0044] FL Federated learning
[0045] Split Learning (SL)
[0046] MTLF Model Training Logical Function
[0047] AnLF Analytics Logical Function
[0048] NWDAF network data analytics function
[0049] UE (User Equipment)
[0050] 3GPP 3G Partnership Program
[0051] OAM Operations, Administrations and Maintenance
[0052] Application Function (AF)
[0053] QoS (Quality of Service)
[0054] NF Network Function
[0055] NG-RAN (Next Generation RAN)
[0056] gNB Next Generation Node B
[0057] BS base station
[0058] Neural Network (NN)
[0059] CNN (Convolutional Neural Network)
[0060] DAG (Directed Acyclic Graph)
[0061] HFL (Hierarchical FL)
[0062] FD Federated Distillation
[0063] FFNN (Feed-Forward Neural Network)
[0064] LSTM (Long Short-Term Memory)
[0065] RNN (Recurrent Neural Network)
[0066] Random Forest (RF)
[0067] EMA (Exponential Moving Average)
[0068] MEC (Multi-access Edge Computing)
[0069] eNA Network Automation Enablers
[0070] CSI Channel State Information
[0071] AP Access Point
[0072] RRC (Radio Resource Control)
[0073] Figure 1This is a schematic diagram of multiple proxy entities 110a-n communicating with a base station 120 and a controller entity 130 in a wireless communication network 100 (e.g., a 5G network 100) according to an embodiment. As used herein, a proxy entity is part of a set of proxy entities 110a-n, wherein each proxy entity 110a-n may be used for: collecting input information, for example, from sensors 112a-c or from channel measurements; processing and compressing the information using, for example, machine learning (ML) models 111a-c, specifically learnable functions of neural networks (NN) 111a-c; and transmitting the processed information to a base station 120 (also referred to herein as access point 120), such as a gNB 120, via a channel of the wireless communication network 100. Each agent entity 110a-n can also receive feedback information from the controller entity 130 via the base station 120, such as actions to be taken by the actuators 113a-c of agent entity 110a-n, or information necessary for performing backpropagation to update the parameters of the ML model 111a-c implemented by each agent entity 110a-n. Furthermore, agent entities 110a-n can perform operations based on action information received from the controller entity 130 via the base station 120, to perform certain actions among all possible actions that the corresponding agent entity 110a-n might take, such as changing its location. The base station or access point 120 is used to collect the outputs of agent entities 110a-n and forward this data to the controller entity 130. Additionally, the base station or access point 120 can receive feedback from the controller entity 130 and forward this feedback to multiple agent entities 110a-n. The controller entity 130 is typically used to process and combine the outputs of multiple ML models 111a-c from multiple agent entities 110a-n, and generate feedback for each agent entity 110a-n on this basis. As will be described in more detail below, the feedback may include information instructing each agent entity 110a-n (e.g., its actuators 113a-c) to take an action or a decompressed version of the output information provided by each agent entity 110a-n. If the agent entities 110a-n and the controller entity 130 are operating in training mode, feedback from the controller entity 130 may be sent back to the agent entities 110a-n via the base station 120 along with information about backpropagation.
[0074] exist Figure 1In the illustrated embodiment, the proxy entities 110a-c implementing ML models 111a-c are UEs 110a-c, and the controller entity 130 is a network entity 130 (in another embodiment, the controller entity 130 may be part of a base station 120). Additional embodiments will be further described below, in which the proxy entities 110a-c and the controller entity 130 are implemented as other types of communication devices, such as mobile robots, drones, micro base stations, etc., for example, implemented as nodes in a 6G network. Other examples include collaborative drones or mobile robots with sensing and communication capabilities, which are considered potential enhancements to the network towards 6G. Such nodes may have sensors and actuators, wherein the actuators move and control the robot's mechanisms, such as moving them in a specific direction, adjusting their transmission power, and activating / deactivating sensing components.
[0075] As mentioned above, Figure 1 Each UE agent entity 110a-c shown is used to run ML model 111a-c, wherein ML model 111a-c is used to process input data into output data with selectable computational complexity and selectable output data size.
[0076] Figure 1 Each UE agent entity 110a-c shown is also used to estimate the current computing resources (also known as computing power) of the agent entity 110a-c and to obtain information indicating the selectable size of the output data of the ML model 111a-c run by the corresponding UE agent entity 110a-c.
[0077] also, Figure 1 Each UE agent entity 110a-c shown is used to select the computational complexity of ML model 111a-c and / or the size of the output data based on the estimation of the computational resources (i.e., computational capabilities) of the corresponding agent entity 110a-c and / or information indicating the selectable size of the output data of ML model 111a-c.
[0078] Therefore, according to the embodiments disclosed herein, agent entities 110a-c and controller entity 130 can adjust the computational level by dynamically adjusting the complexity of ML models 111a-c at runtime. In other embodiments, agent entities 110a-c can adjust for different levels of communication resources (i.e., communication capabilities) by dynamically adjusting the compression of the output of each ML model 111a-c of each agent entity 110a-c. In other words, the embodiments disclosed herein enable the learning process to adapt to current communication and computational resources / capabilities at runtime.
[0079] like Figure 1 As shown, in one embodiment, Figure 1The interaction between the UE agent entities 110a-c and the controller entity 130 can be achieved as follows: For each agent entity 110a-c, a minimum ML model 111a-c with respect to complexity and compression level can be fixed. These minimum ML models 111a-c are trained until the system can no longer learn more at the set complexity and compression level. At this stage, the weights of the ML models 111a-c can be fixed or frozen, and more neurons or processing layers can be added to the ML models 111a-c to increase complexity and reduce compression. These enhanced ML models 111a-c are retrained until the system can no longer learn more at the set complexity and compression level. The previous tuning and training steps are repeated until all desired complexity and compression levels are trained. At the controller entity 130, the current compression and complexity levels are collected, which can be used for post-processing and decompression, respectively. During deployment, the agent entities 110a-c send the complexity and compression levels to the controller entity 130 via the base station 120.
[0080] Therefore, the embodiments disclosed herein may involve one or more of the following features: base station 120 sends network conditions to agent entities 110a-c and controller entity 130; according to, for example Figure 2 The table shown maps node conditions such as bps, processing capacity, and latency to execution policies; the selected execution policy is sent from agent entities 110a-c to controller entity 130 via base station 120; the complexity and compression level of agent entities 110a-c and controller entity 130 are dynamically adjusted. In one embodiment, Figure 2 The complexity and compression index shown in the table indicate the percentage of all layers of the corresponding ML model 110a-c used to achieve a specific complexity and compression level of ML model 110a-c.
[0081] Figure 3 It is shown Figure 1 Dynamically adaptive signaling diagrams for each agent entity 110a-c in ML model 111a-c.
[0082] exist Figure 3 In step 0, the controller entity 130 and multiple UE agent entities 110a-n exchange execution policy mappings through the base station 120, for example... Figure 2 The execution strategy table is shown.
[0083] exist Figure 3 In step 1, base station 120 shares information about network conditions with multiple UE agent entities 110a-n and controller entity 130.
[0084] exist Figure 3In step 2, each UE agent entity 110a-n is configured according to a complexity index and compression level, for example, based on... Figure 2 The table shown is used to select the ML model execution strategy.
[0085] exist Figure 3 In step 3, each UE agent entity 110a-n determines the output of ML model 110a-c based on the execution strategy selected in the previous step.
[0086] exist Figure 3 In step 4, each UE agent entity 110a-n sends the output of ML model 110a-c and the execution strategy to the controller entity 130.
[0087] exist Figure 3 In steps 5 and 6, the controller entity 130 determines the action of the UE agent entity 110a-n based on the output of the ML model 111a-c and the execution strategy from the UE agent entity 110a-n, and may also based on other conditions of the controller entity 130.
[0088] exist Figure 3 In step 7, the controller entity 130 will update the action and ML model parameters determined in the previous step and feed them back to the UE agent entity 110a-n.
[0089] exist Figure 3 In step 8, each UE agent entity 110a-n can perform actions and update its ML model parameters based on feedback received from controller entity 130.
[0090] Therefore, in the embodiments, each agent entity 110a-n can be used to perform the following operations:
[0091] 1. Collect current communication conditions (e.g., bps, latency) and computing power.
[0092] 2. Look up the mapping table from the complexity index and compression level (e.g., Figure 2 The execution strategy in the execution strategy table shown .
[0093] 3. According to the execution strategy Calculate the output from the input. .
[0094] 4. Send output to controller entity 130 via base station 120 and execution strategy .
[0095] In this embodiment, the controller entity 130 is further configured to perform the following operations:
[0096] 5. Collect all outputs from proxy entities 110a-n and execution strategy .
[0097] 6. According to the execution strategy For all outputs The system performs decompression and processing, and generates output based on controller conditions and system status. Optionally, the gradient is calculated and the controller parameters are updated (if during training).
[0098] 7. Optional, feedback Output to agent entity 110a-n and gradients (if during training).
[0099] All agent entities 110a-c receive output from controller entity 130. And accordingly execute and update model 111a-c (if during training).
[0100] Other embodiments of the agent entity and controller entity will be described below.
[0101] The first other embodiment relates to the compression of channel state information (CSI) in MIMO FDD systems. As will be understood, CSI information is used for transmission parameter decisions, such as selecting modulation and coding schemes, transmission layers, etc., which are necessary to achieve the desired communication system performance. This is primarily accomplished by relying on pilots sent from the transmitter to the receiver, which then shares the estimated channel information or associated channel parameters back to the transmitter. With the increasing number of transmit and receive antennas, CSI feedback information can consume a significant amount of uplink bandwidth. To address the increased bandwidth requirements of sharing CSI feedback, the embodiments disclosed herein implement efficient sharing of CSI information derived from a reference signal, such as CSI-RS, by taking into account the communication resource conditions (e.g., data rate, latency, etc.) and computational resource conditions (i.e., capabilities, storage capacity) of the nodes involved. Current schemes in 3GPP can share parameters such as RI, PMI, CQI, etc., derived from CSI reporting parameters and predefined mechanisms (e.g., existing codebooks).
[0102] According to the embodiments, for existing mechanisms and other potential flexible transmission adaptation mechanisms that may rely on the original channel estimation (e.g., the channel matrix derived from the reference signal), each agent entity 110a-n is capable of compressing the CSI feedback information. More specifically, each agent entity 110a-n is used to share and process compressed CSI feedback information by dynamically changing the compression level according to the communication resource conditions and computing resources at the respective node.
[0103] Figure 4 The signaling diagram illustrates the first scenario of compressed downlink CSI transmission in the uplink (including...). Figure 4 (Steps 1 to 7 shown). In this case, the compressed CSI information is shared from the transmitter, which is considered a UE agent entity 110a, to the receiver, which is considered a base station 120. This corresponds to the transmission of compressed downlink CSI to support transmission adaptation at base station 120. Controller entity 130 is part of base station 120. For example, an autoencoder is considered to compress and decompress CSI information at UE agent entity 110a and base station 120, respectively. More specifically, UE agent entity 110a hosts the encoder for compressing CSI information and transmitting it over the air interface. Base station 120, including controller entity 130, hosts the decoder for decompressing it upon reception according to the execution policy index used. In this embodiment, it is assumed that an autoencoder model with only one agent entity is trained and deployed at base stations 120, 130, and UE 110a. The compression configuration of the autoencoder can be based on network conditions at base station 120, such as the channel quality of all users connected to base station 120 or the load at base station 120. Therefore, base station 120 can determine the compression level and share the compression level with UE 110a. Base stations can also be used to share the mapping between available communication and computing resources and execution policy indexes.
[0104] After receiving the compression level from base station 120, UE 110a determines the complexity level (of the compression / decompression process) based on its computing power (e.g., depending on battery status) and a shared table, and therefore from the shared execution strategy table, such as... Figure 2 The table shown determines the execution strategy. CSI feedback at UE 110a can be compressed based on this decision. UE 110a can share the compressed output and associated execution strategy as part of the CSI report (see [link to CSI report]). Figure 4 Step 6) is to support decompression at the base station.
[0105] The embodiments described above can be implemented in current communication systems by enhancing RRC cells. In one embodiment, a CSI-related RRC cell (e.g., CSI-ReportConfig) may include the following elements:
[0106] ●The new id in reportConfigId – used to indicate autoencoder-based CSI feedback for downlink CSI;
[0107] ●New quantities in reportQuantity – Compressed downlink CSI, execution strategy;
[0108] ● New field in CSImodel – for sharing the encoder portion of an autoencoder-based CSI feedback model, for sharing the mapping from communication and computational resources to execution strategies;
[0109] ● New field for CSImodel parameters – Retrieves the physical location of the compression level (i.e., the location of the physical resource unit, for example, in the DCI field of PDCCH).
[0110] The transmission of CSI reports from UE 110a can be performed within the PUCCH. As described above, CSI reports can be extended with coded channel information and execution strategies (see [link to documentation]). Figure 4 Step 6).
[0111] Figure 5 The signaling diagram illustrates a second scenario regarding the transmission of compressed uplink CSI in the downlink (including...). Figure 5 (Steps 1 to 6 shown). In this case, compressed CSI information is shared from base station 120 to UE 110a. This corresponds to the transmission of compressed uplink CSI to support transmission adaptation at UE 110a. In this embodiment, the controller entity is part of UE 110a. Since base station 120 shares compressed CSI feedback information, the base station knows the compression level and computation level to be adopted, and therefore can, for example, according to... Figure 2 The execution policy table shown selects the execution policy locally. This execution policy and compressed CSI output are shared with UE 110a (see [link]). Figure 5 Step 1).
[0112] Similar to the first scenario described above, this second scenario can be achieved in current communication systems by enhancing RRC cells. For example, a CSI-related RRC cell (e.g., CSI-ReportConfig) may include the following elements:
[0113] ●The new id in reportConfigId – used to indicate autoencoder-based CSI feedback for uplink CSI;
[0114] ●New quantities in reportQuantity – Compressed uplink CSI, execution strategy;
[0115] ● New field in CSImodel – for sharing the encoder portion of an autoencoder-based CSI feedback model, for sharing the mapping from communication and computational resources to execution strategies.
[0116] Encoded channel information and execution strategies can be used to enhance the transmission of CSI reports from base station 120 on the PDCCH (see [link]). Figure 5 Step 5).
[0117] Figure 6 Message exchange is illustrated in another embodiment, where multiple agent entities are collaborative drones implementing corresponding micro base stations to enhance network coverage (e.g., for critical V2X applications). In this embodiment, controller entity 130 is implemented as RANDAF / BS controller entity 130, and agent entities 110a-n are implemented as drone micro base stations 110a-n providing enhanced coverage to network users. Input (sensor) information for each drone agent entity 110a-n may include its location, i.e., its x, y, z coordinates; the direction / angle and transmit power of each antenna used by the drone agent entity 110a-n to cover the area below; the number of users supported by each of these antennas; the average uplink and downlink traffic load and throughput of each antenna; and an estimate of the overlap of each antenna's coverage area with that of neighboring drones. Feedback actions from controller entity 130 may include the next (target) location of the corresponding drone agent entity 110a-n, and the direction, angle, and transmit power of each antenna used by the corresponding drone agent entity 110a-n.
[0118] exist Figure 6 In step 1, gNB 120, for example, shares information about network conditions with multiple drone agent entities 110a-n and controller entity 130 via the Xn-C interface.
[0119] exist Figure 6 In step 2, each drone agent entity 110a-n, for example, according to Figure 2 The table shown selects the ML model execution strategy based on the complexity level and compression function.
[0120] exist Figure 6 In step 3, each drone agent entity 110a-n determines the output of the ML model based on the execution strategy selected in the previous step.
[0121] exist Figure 6 In step 4, each drone agent entity 110a-n sends the output of the ML model and the execution strategy to the controller entity 130.
[0122] exist Figure 6 In steps 5 and 6, the controller entity 130 determines the actions of the drone agent entities 110a-n based on the output of the ML models 111a-c and the execution strategies from the drone agent entities 110a-n, and possibly based on other conditions of the controller entity 130.
[0123] exist Figure 6 In step 7, the controller entity 130 will feed back the actions determined in the previous step to the drone agent entities 110a-n.
[0124] exist Figure 6 In step 8, each drone agent entity 110a-n can perform actions, such as changing its position, and update its ML model parameters based on the actions received from the controller entity 130.
[0125] In another embodiment, the agent entity may be a mobile robot agent entity used in a factory to provide sensing and communication capabilities to the machine. The difference from the previous embodiments is that the input information may also include feedback to the mobile robot, such as actions it intends to take, or requests for new types of sensing information submitted by the mobile robot. Therefore, in addition to what has been described with respect to the previous embodiments, in this embodiment, feedback actions from the controller entity may also include requests for the mobile robot agent entity to activate new sensing components or deactivate unused sensing components of the mobile robot agent entity, for example, to conserve the energy of the mobile robot agent entity.
[0126] Figure 7 This is a flowchart illustrating a method 700 for operating proxy entities such as UE proxy entities 110a-n or base station proxy entity 120 for adaptive learning. Method 700 includes a step 701 of running a machine learning (ML) model (e.g., ML model 111a-c), wherein the ML model 111a-c processes input data into output data with a selectable computational complexity and a selectable output data size. Furthermore, method 700 includes a step 703 of estimating the computational resources of proxy entities 110a-n; 120 and a step 705 of obtaining information indicating a selectable size of the output data of ML model 111a-c. Method 700 also includes a step 707 of selecting the computational complexity and / or the size of the output data of ML model 111a-c based on the estimation of the computational resources of proxy entities 110a-n; 120 and / or the information indicating a selectable size of the output data of ML model 111a-c.
[0127] According to an embodiment, method 700 can be performed by each UE proxy entity 110a-n or base station proxy entity 120. Therefore, other features of method 700 are directly derived from the functionality of UE proxy entities 110a-n and base station proxy entity 120 and their different embodiments described above and below.
[0128] As will be understood, the embodiments disclosed herein can dynamically adapt to the complexity and, for example, compression level of the ML model of the agent entities in a segmentation learning environment. This enables each agent entity to adapt to a dynamic wireless environment and saves storage space for the adjusted ML model. The efficient selection of the CSI compression level implemented by the embodiments disclosed herein can dynamically adjust the data rate of the control channel based on channel conditions and the computational capacity of each agent entity (e.g., depending on the battery state of the respective agent entity). Furthermore, the embodiments disclosed herein enable cooperation between robotic agent entities in dynamic environments, as well as coordination between coupled BSs (e.g., macro base stations) and micro / femto base stations.
[0129] Those skilled in the art will understand that “blocks” (“units”) in the various drawings (methods and apparatuses) represent or describe the functionality of embodiments of this disclosure (and are not necessarily independent “units” in hardware or software), thereby equally describing the functionality or features (units equivalent to steps) of apparatus embodiments and method embodiments.
[0130] In the various embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the described apparatus embodiments are merely exemplary. For example, the unit division is only a logical functional division, and other division methods may be used in actual implementation. For example, multiple units or components may be combined or integrated into another system, or some features may be omitted or not performed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed can be implemented through some interfaces. Indirect coupling or communication connection between devices or units can be implemented electronically, mechanically, or otherwise.
[0131] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of the embodiment solution according to actual needs.
[0132] Furthermore, the functional units in the embodiments of this disclosure can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.
Claims
1. A proxy entity (110a-n; 120) for adaptive learning, characterized in that, The proxy entity (110a-n; 120) is used for: Run machine learning (ML) models (111a-c), wherein the ML models (111a-c) are used to process input data into output data with selectable computational complexity and selectable output data size; Estimate the computing resources of the agent entities (110a-n; 120); Obtain information indicating the selectable size of the output data of the ML model (111a-c); Based on the information regarding the estimated computing resources of the agent entity (110a-n; 120) and / or the selectable size of the output data of the ML model (111a-c), the computational complexity of the ML model (111a-c) and / or the size of the output data are selected.
2. The proxy entity (110a-n; 120) according to claim 1, characterized in that, The agent entity (110a-n; 120) is used to receive, via a wireless communication channel, the information indicating the selectable size of the output data of the ML model (111a-c) from the controller entity (130; 110a).
3. The proxy entity (110a-n; 120) according to claim 1, characterized in that, In order to obtain the information indicating the selectable size of the output data of the ML model (111a-c), the proxy entity (110a-n; 120) is used to estimate the communication resources for communicating with the controller entity (130; 110a) via a wireless communication channel, wherein the proxy entity (110a-n; 120) is used to select the computational complexity of the ML model (111a-c) and / or the size of the output data based on the estimation of the computational resources and the estimation of the communication resources.
4. The proxy entity (110a-n; 120) according to claim 3, characterized in that, In order to estimate the communication resources, the agent entity (110a-n; 120) is used to obtain information indicating the current data rate for communication between the agent entity (110a-n; 120) and the controller entity (130; 110a) via the wireless communication channel, wherein the agent entity (110a-n; 120) is used to select the computational complexity of the ML model (111a-c) and / or the size of the output data based on the estimation of the agent entity's (110a-n; 120) computational resources and the information indicating the current data rate.
5. The agent entity (110a-n; 120) according to any one of claims 2 to 4, characterized in that, The agent entity (110a-n; 120) is also used to send the output data of the ML model to the controller entity (130; 110a) via the wireless communication channel.
6. The proxy entity (110a-n; 120) according to claim 5, characterized in that, In response to sending the output data of the ML model (111a-c) to the controller entity (130; 110a), the agent entity (110a-n; 120) is further configured to receive response data from the controller entity (130; 110a), wherein the response data is based on the output data of the ML model (111a-c) of the agent entity (110a-n; 120) and multiple other output data of multiple other ML models (111a-c) of multiple other agent entities (110a-n; 120).
7. The proxy entity (110a-n; 120) according to claim 6, characterized in that, The response data includes information instructing the agent entity (110a-n; 120) to take action and / or information for performing backpropagation to update the ML model (111a-c).
8. The agent entity (110a-n; 120) according to any one of claims 2 to 7, characterized in that, The proxy entity (110a-n) is a user equipment (110a-n) used to exchange data with the controller entity (130) through the wireless communication channel and the base station (120).
9. The proxy entity (110a-n) according to claim 8, characterized in that, The ML model (111a-c) is the encoding part of an autoencoder, wherein the input data of the encoding part of the autoencoder is channel state information, and the output data of the encoding part of the autoencoder is compressed channel state information.
10. The agent entity (110a-n; 120) according to any one of claims 2 to 7, characterized in that, The agent entity (120) is a mobile micro base station (120) among a plurality of mobile micro base stations, wherein the output data of each mobile micro base station (120) enables the controller entity (130) to coordinate the plurality of mobile micro base stations (120).
11. The agent entity (120) according to any one of claims 2 to 7, characterized in that, The agent entity (120) is a base station (120), and the controller entity (110a) is a user equipment (110a).
12. The proxy entity (110a-n; 120) according to any one of the preceding claims, characterized in that, The ML model (111a-c) includes multiple processing layers for processing the input data into the output data. In order to select the computational complexity of the ML model (111a-c), the proxy entity (110a-n; 120) is used to select a selectable number of processing layers from the multiple processing layers of the ML model (111a-c).
13. The proxy entity (110a-n; 120) according to any one of the preceding claims, characterized in that, The proxy entity (110a-n; 120) includes a battery for powering one or more processors of the proxy entity (110a-n; 120) to implement the ML model (111a-c), wherein the proxy entity (110a-n; 120) is used to estimate the computing resources of the proxy entity (110a-n; 120) based on the state of charge of the battery.
14. A method (700) for manipulating agent entities (110a-n; 120) for adaptive learning, characterized in that, The method (700) includes: Run (701) machine learning (ML) models (111a-c), wherein the ML models (111a-c) are used to process input data into output data with selectable computational complexity and selectable output data size; Estimate the computational resources of the proxy entity (110a-n; 120) described in (703); Obtain (705) information indicating the selectable size of the output data of the ML model (111a-c); Based on the information regarding the estimated computational resources of the agent entity (110a-n; 120) and / or the selectable size of the output data of the ML model (111a-c), the computational complexity of the ML model (111a-c) and / or the size of the output data are selected (707).
15. A computer program product, characterized in that, Includes a computer-readable storage medium for storing program code, which, when executed by a computer or processor, causes the computer or processor to perform the method (700) according to claim 14.