Data processing method and apparatus, and related device

Abstract

The application discloses a data processing method and device and related equipment thereof, and relates to the field of wireless communication. The specific implementation scheme is as follows: initial state information of a plurality of computing resource nodes is acquired; a target computing power allocation strategy corresponding to at least one target computing resource node in the plurality of computing resource nodes is determined according to the initial state information of the plurality of computing resource nodes; and the target computing power allocation strategy is sent to the at least one target computing resource node, wherein the target computing power allocation strategy comprises target computing power allocation ratios of the plurality of computing resource nodes, and is used for the corresponding target computing resource node to distribute to-be-processed data according to the target computing power allocation ratios of the plurality of computing resource nodes. Therefore, the amount of to-be-processed data of each computing resource node can be dynamically adjusted, the amount of data processed by each computing resource node is matched with the computing power thereof, the resource utilization rate of each computing resource node in a mobile communication system is improved, and the data processing efficiency in the mobile communication system is improved.

Description

Technical Field

[0001] This application relates to the field of wireless communication technology, and in particular to a data processing method, apparatus and related equipment. Background Technology

[0002] In mobile communication systems, machine learning is introduced to replace traditional communication modules or assist system decision-making, thereby improving the performance of the communication system. The online training process for these machine learning applications requires different computing resource nodes to process data from different sources and execute a large number of computational tasks. These computing resource nodes can include terminals, base stations, core network equipment, and edge computing devices.

[0003] However, due to limitations in the computing capabilities of different computing resource nodes, when the amount of data to be processed on a computing resource node is large, it may cause the computing resource node to be unable to complete the computing task in a timely manner, resulting in business delays and potentially affecting the overall performance of the node. Summary of the Invention

[0004] This application provides a data processing method, apparatus, and related equipment.

[0005] According to one aspect of this application, a data processing method is provided, the method being executed by a computing center node, the method comprising:

[0006] Obtain the initial status information of multiple computing resource nodes;

[0007] Based on the initial state information of the plurality of computing resource nodes, determine the target computing power allocation strategy corresponding to at least one target computing resource node among the plurality of computing resource nodes;

[0008] A corresponding target computing power allocation strategy is sent to the at least one target computing resource node, wherein the target computing power allocation strategy includes the target computing power allocation ratio of the plurality of computing resource nodes, and is used by the corresponding target computing resource node to distribute the data to be processed according to the target computing power allocation ratio of the plurality of computing resource nodes.

[0009] Optionally, the initial state information includes at least one of the following: the amount of data to be processed for the corresponding computing resource node, the amount of computation required for the corresponding computing resource node to process the data to be processed, the available storage space of the corresponding computing resource node, and the available computing power of the corresponding computing resource node.

[0010] The available computing power is used to characterize the amount of data that the corresponding computing resource node can process per unit time.

[0011] Optionally, before obtaining the initial state information of multiple computing resource nodes, the method further includes:

[0012] Obtain at least one of the following: the location of each computing resource node, the data characteristics of the data to be processed by each computing resource node, and the data type.

[0013] The plurality of computing resource nodes are determined from the computing resource nodes based on at least one of the location of each computing resource node, the data characteristics of the data to be processed by each computing resource node, and the data type.

[0014] Send indication information to the plurality of computing resource nodes, wherein the indication information is used to instruct the plurality of computing resource nodes to send corresponding initial status information to the computing center node.

[0015] Optionally, determining the target computing power allocation strategy corresponding to at least one target computing resource node among the plurality of computing resource nodes based on the initial state information of the plurality of computing resource nodes includes:

[0016] A state sequence is generated based on the initial state information of the multiple computing resource nodes;

[0017] An action sequence is generated based on the initial state information of the plurality of computing resource nodes and the initial computing power allocation strategy of the at least one target computing resource node; wherein, the initial computing power allocation strategy includes the initial computing power allocation ratio of the plurality of computing resource nodes;

[0018] Based on the state sequence and the action sequence, a cumulative reward is determined; wherein the cumulative reward is inversely related to the amount of resources or time required to execute the action sequence in the state indicated by the state sequence.

[0019] Based on the accumulated reward, the initial computing power allocation strategy of the at least one target computing resource node is updated at least once, and a reinforcement learning model is trained to obtain the target computing power allocation strategy corresponding to the at least one target computing resource node.

[0020] Optionally, the reinforcement learning model is trained through the following steps:

[0021] Obtain training samples, which include the state sequence of the plurality of computing resource nodes, the action sequence corresponding to the maximum cumulative reward, the reward obtained after taking the action sequence corresponding to the maximum cumulative reward under the state sequence, and the state sequence after taking the action sequence corresponding to the maximum cumulative reward, wherein the action sequence corresponding to the maximum cumulative reward is determined according to the computing power allocation ratio of the plurality of computing resource nodes that maximizes the cumulative reward;

[0022] A reinforcement learning model is trained based on the training samples to update the model parameters of the reinforcement learning model;

[0023] If the model training termination condition is not met, the updated reinforcement learning model is trained using the updated training samples until the model training termination condition is met, at which point the training of the reinforcement learning model is stopped.

[0024] Optionally, the model training termination condition includes at least one of the following conditions: the cumulative reward of the reinforcement learning model converges, the training time of the reinforcement learning model reaches a preset threshold, and the number of iterations of the reinforcement learning model reaches a set number.

[0025] Optionally, the method further includes:

[0026] Receive the data to be processed sent by the plurality of computing resource nodes;

[0027] The target model is trained based on the data to be processed from the multiple computing resource nodes to obtain the target model after training by the computing center node.

[0028] Receive the target model trained by the computing resource nodes sent by the plurality of computing resource nodes;

[0029] The target model trained by the computing center node and the target models trained by each computing resource node are aggregated to obtain the aggregated target model.

[0030] The aggregated target model is sent to the multiple computing resource nodes.

[0031] According to another aspect of this application, another data processing method is provided, the method being executed by a target computing resource node, the method comprising:

[0032] Send the initial status information of the target computing resource node to the computing center node;

[0033] The system receives a target computing power allocation strategy sent by the computing center node, wherein the target computing power allocation strategy includes a target computing power allocation ratio for multiple computing resource nodes.

[0034] The data to be processed on the target computing resource nodes is distributed according to the target computing power allocation ratio of the multiple computing resource nodes.

[0035] Optionally, before sending the initial state information of the target computing resource node to the computing center node, the method further includes:

[0036] The system receives an instruction message sent by the computing center node, wherein the instruction message is used to instruct the corresponding initial state information to be sent to the computing center node.

[0037] Optionally, distributing the data to be processed to the target computing resource nodes according to the target computing power allocation ratio of the plurality of computing resource nodes includes:

[0038] From the unprocessed data of the target computing resource node, determine the first target data that matches the target computing power allocation ratio corresponding to the target computing resource node;

[0039] Based on the target computing power allocation ratio corresponding to the parent computing resource node of the target computing resource node among the plurality of computing resource nodes, the remaining data in the unprocessed data of the target computing resource node, excluding the first target data, is sent to the parent computing resource node.

[0040] Optionally, the method further includes:

[0041] The system receives second target data sent by each subordinate computing resource node corresponding to the target computing resource node, wherein the second target data is determined based on the target computing power allocation ratio corresponding to each subordinate computing resource node and the data to be processed corresponding to each subordinate computing resource node.

[0042] Optionally, the method further includes:

[0043] The target model is trained based on the second target data and the first target data sent by each of the lower-level computing resource nodes to obtain the target model trained by the target computing resource nodes.

[0044] Send the target model, trained by the target computing resource node, to the computing center node;

[0045] Receive the aggregated target model sent by the computing center node.

[0046] According to another aspect of this application, a computing center node is provided, comprising:

[0047] Memory, transceiver, processor:

[0048] A memory for storing computer programs; a transceiver for sending and receiving data under the control of the processor; and a processor for reading the computer programs from the memory and performing the following operations:

[0049] Obtain the initial status information of multiple computing resource nodes;

[0050] Based on the initial state information of the plurality of computing resource nodes, determine the target computing power allocation strategy corresponding to at least one target computing resource node among the plurality of computing resource nodes;

[0051] A corresponding target computing power allocation strategy is sent to the at least one target computing resource node, wherein the target computing power allocation strategy includes the target computing power allocation ratio of the plurality of computing resource nodes, and is used by the corresponding target computing resource node to distribute the data to be processed according to the target computing power allocation ratio of the plurality of computing resource nodes.

[0052] According to another aspect of this application, a target computing resource node is provided, comprising:

[0053] Memory, transceiver, processor:

[0054] A memory for storing computer programs; a transceiver for sending and receiving data under the control of the processor; and a processor for reading the computer programs from the memory and performing the following operations:

[0055] Send the initial status information of the target computing resource node to the computing center node;

[0056] The system receives a target computing power allocation strategy sent by the computing center node, wherein the target computing power allocation strategy includes a target computing power allocation ratio for multiple computing resource nodes.

[0057] The data to be processed on the target computing resource nodes is distributed according to the target computing power allocation ratio of the multiple computing resource nodes.

[0058] According to another aspect of this application, a data processing apparatus is provided, comprising:

[0059] The acquisition unit is used to acquire the initial state information of multiple computing resource nodes;

[0060] The determining unit is used to determine the target computing power allocation strategy corresponding to at least one target computing resource node among the plurality of computing resource nodes based on the initial state information of the plurality of computing resource nodes.

[0061] The sending unit is configured to send a corresponding target computing power allocation strategy to the at least one target computing resource node, wherein the target computing power allocation strategy includes the target computing power allocation ratio of the plurality of computing resource nodes, and the corresponding target computing resource node distributes the data to be processed according to the target computing power allocation ratio of the plurality of computing resource nodes.

[0062] According to another aspect of this application, another data processing apparatus is provided, comprising:

[0063] The sending unit is used to send the initial status information of the target computing resource node to the computing center node;

[0064] A receiving unit is configured to receive a target computing power allocation strategy sent by the computing center node, wherein the target computing power allocation strategy includes a target computing power allocation ratio for multiple computing resource nodes.

[0065] The distribution unit is used to distribute the data to be processed by the target computing resource nodes according to the target computing power allocation ratio of the multiple computing resource nodes.

[0066] According to another aspect of this application, a processor-readable storage medium is provided, characterized in that the processor-readable storage medium stores a computer program for causing the processor to perform the data processing method described in any of the above embodiments.

[0067] According to another aspect of this application, a computer program product is provided, including a computer program that, when executed by a processor, implements the data processing method described in any of the above embodiments of this application.

[0068] The data processing method, apparatus, and related equipment provided in this application embodiment acquire initial state information of multiple computing resource nodes; determine a target computing power allocation strategy corresponding to at least one target computing resource node among the multiple computing resource nodes based on the initial state information of the multiple computing resource nodes; and send the corresponding target computing power allocation strategy to the at least one target computing resource node. The target computing power allocation strategy includes a target computing power allocation ratio for the multiple computing resource nodes, which is used by the corresponding target computing resource node to distribute data to be processed according to the target computing power allocation ratio of the multiple computing resource nodes. Therefore, it is possible to generate a computing power allocation strategy corresponding to each computing resource node based on the state of each computing resource node, so that each computing resource node distributes the corresponding data to be processed according to the corresponding computing power allocation strategy. This allows for dynamic adjustment of the amount of data to be processed by each computing resource node, ensuring that the amount of data processed by each computing resource node matches its own computing power, thereby improving the resource utilization of each computing resource node in the mobile communication system and improving the data processing efficiency of the mobile communication system.

[0069] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this application, nor is it intended to limit the scope of this application. Other features of this application will become readily apparent from the following description. Attached Figure Description

[0070] The accompanying drawings are provided for a better understanding of this solution and do not constitute a limitation of this application. Wherein:

[0071] Figure 1This is a flowchart illustrating the data processing method provided in Embodiment 1 of this application;

[0072] Figure 2 This is a schematic diagram of the computing power allocation system in an embodiment of this application;

[0073] Figure 3 A flowchart illustrating another data processing method provided in an embodiment of this application;

[0074] Figure 4 A flowchart illustrating another data processing method provided in an embodiment of this application;

[0075] Figure 5 A flowchart illustrating another data processing method provided in an embodiment of this application;

[0076] Figure 6 This is a schematic diagram of the learning process for the computing power allocation strategy in the embodiments of this application;

[0077] Figure 7 This is a schematic diagram of the structure of a self-encoder;

[0078] Figure 8 This is a schematic diagram of the CSI feedback and channel reconstruction system in an embodiment of this application;

[0079] Figure 9 This is a schematic diagram of the learning process for CSI feedback and channel reconstruction computing power allocation strategy in an embodiment of this application;

[0080] Figure 10 A flowchart illustrating another data processing method provided in an embodiment of this application;

[0081] Figure 11 This is a schematic diagram of the structure of a computing center node provided in an embodiment of this application;

[0082] Figure 12 This is a schematic diagram of the structure of a target computing resource node provided in an embodiment of this application;

[0083] Figure 13 This is a schematic diagram of the structure of a data processing apparatus provided in an embodiment of this application;

[0084] Figure 14 This is a schematic diagram of another data processing apparatus provided in an embodiment of this application. Detailed Implementation

[0085] In the embodiments of this application, the term "and / or" describes the relationship between associated objects, indicating that three relationships can exist. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. The character " / " generally indicates that the preceding and following associated objects have an "or" relationship.

[0086] In the embodiments of this application, the term "multiple" refers to two or more, and other quantifiers are similar.

[0087] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0088] In mobile communication systems, introducing machine learning to replace traditional communication modules or assist system decision-making requires first generating an Artificial Intelligence (AI) model through a model training process, then deploying or updating the model to execute inference services, and finally optimizing the system based on the inference results.

[0089] The model training process involves computing resource nodes such as terminals, base stations, core network equipment, and edge computing devices. During operation, each computing resource node generates or collects massive amounts of data and processes and performs related calculations on this data. Each computing resource node can execute model training tasks in a separate offline training system, or it can complete model training online within the mobile communication system.

[0090] It should be understood that performing online training tasks in mobile communication systems allows for the immediate acquisition of real-world data from the existing network, enabling the timely generation of training models that are more closely aligned with application scenarios, thus ensuring the adaptability of policy updates and the accuracy of decision support.

[0091] However, when performing online training tasks in mobile communication systems, the nodes that generate training data are often not the same as the nodes with powerful computing capabilities. The computing power of the nodes generating training data is often insufficient to meet the computational demands of online training, potentially causing the nodes to be unable to complete their computational tasks in a timely manner. In other words, when the data generated by a node and its computing power are mismatched, it is inevitable to solve the problems of data transmission and training task allocation.

[0092] The first approach is to process data locally, performing relevant training tasks at the data generation / collection node (e.g., a terminal or Internet of Things (IoT) sensor node).

[0093] The second approach is to send all data to a specific aggregation node (e.g., a base station or edge computing device) for unified processing and to perform comprehensive training.

[0094] The third approach is to combine the two methods mentioned above, with training occurring partly at the data generation / collection nodes and partly at specific aggregation nodes, and with data and model parameters interacting between the data generation / collection nodes and the specific aggregation nodes.

[0095] However, in the first approach described above, when the data volume is large and / or the data generation rate is high, the computing power of the data generation / collection nodes may be insufficient. This can lead to the data generation / collection nodes being unable to complete their computing tasks in a timely manner, resulting in service delays and potentially impacting the overall performance of the nodes. For example, when the data generation / collection node is a terminal, if the terminal needs to process a large amount of data or perform heavy training tasks, it may affect the terminal's call and data services, and may also consume more power, leading to a reduced usability experience for the terminal.

[0096] In the second approach described above, all data must be sent to a specific aggregation node via the network, where unified training is then performed. When training data is generated frequently, transmitting large amounts of training data over the network can cause instantaneous peaks in network load, easily leading to network congestion. Furthermore, excessive communication overhead may affect the performance of other services in the mobile communication system. In addition, the performance of the aggregation node may become a bottleneck for the overall training task, limiting the availability of the online training task to the hardware configuration of the aggregation node.

[0097] In the third approach described above, there are no clear criteria for determining which data to use for training at the data generation / collection nodes and which data to send to the aggregation nodes for training, making it difficult to control the splitting scale. While hybrid training can address the data-computing mismatch to some extent, it is difficult to control and increases the complexity of implementing the training algorithm.

[0098] Therefore, in order to solve the above problems, this application provides a data processing method, apparatus and related equipment. The method and apparatus are based on the same application concept. Since the methods and apparatus solve problems in similar principles, the implementation of the apparatus and method can refer to each other, and repeated parts will not be described again.

[0099] The data processing method, apparatus, and related equipment provided in this application will be described in detail below with reference to the accompanying drawings.

[0100] Figure 1 This is a schematic flowchart of a data processing method provided in an embodiment of this application.

[0101] The data processing method of this application embodiment can be executed by a computing center node. Here, the computing center node refers to the control node in a mobile communication system used to generate computing power allocation strategies.

[0102] As an example, this section illustrates the application of multiple computing resource nodes in an online training scenario. The computing center node can be an edge computing node, or it can be a core node. For instance, if the computing resource nodes (such as data generation / collection nodes) are end nodes or aggregation nodes, the computing center node can be a pre-defined edge computing node to be as close as possible to the data side and improve computing efficiency; if the computing resource nodes (such as data generation / collection nodes) are core nodes, the computing center node can be a pre-defined core node.

[0103] The explanation of each node can be found in the following embodiments, and will not be repeated here.

[0104] like Figure 1 As shown, the data processing method may include the following steps:

[0105] Step 101: Obtain the initial status information of multiple computing resource nodes.

[0106] In this embodiment of the application, computing resource nodes may include, but are not limited to, terminals, base stations, core network equipment, and multi-access edge computing (MEC) devices.

[0107] As an example, the application of multiple computing resource nodes in an online training scenario is used to illustrate the concept. Based on the role or purpose of computing resource nodes in a mobile communication system, computing resource nodes can be divided into data generation / collection nodes, general computing nodes, intermediate nodes, computing center nodes, and strategy libraries.

[0108] Among them, data generation / collection nodes are nodes in mobile communication systems used to generate or collect data (such as training data).

[0109] General computing nodes: These are the computing nodes that perform collaborative training functions.

[0110] Intermediate nodes are computing nodes that do not perform collaborative training functions. That is, the data generation / collection nodes do not offload the online training tasks to the intermediate nodes, but they still need to communicate with the computing center nodes through the intermediate nodes.

[0111] The computing center node is the control node that generates the computing power allocation strategy.

[0112] Strategy library: This is a database that stores computing power allocation strategies in mobile communication systems.

[0113] As another example, based on the location and function of computing resource nodes in the network topology of a mobile communication system, computing resource nodes can be further divided into end nodes, aggregation nodes, edge computing connection nodes, edge computing nodes, and core nodes.

[0114] End nodes are the computing nodes closest to the user, and can be terminals, IoT devices, etc., possessing basic computing capabilities. If the end node is a terminal or mobile device, it is generally battery-powered and is relatively sensitive to energy consumption; if the end node is an IoT device, it may be powered by a power supply or battery. End nodes typically function as data generation / collection nodes, but may also function as general computing nodes when privacy is not a concern and power supply is sufficient.

[0115] Aggregation node: As the superior node of the end node, it can be a base station or a data collection and aggregation node in the Internet of Things (IoT). Among them, the aggregation node can serve as a data generation / collection node, an intermediate node, or a general computing node.

[0116] Edge computing nodes typically possess greater computing power than aggregation nodes and can be MEC devices or application (App) servers located close to the radio access network. Edge computing nodes connect to aggregation and core nodes via edge computing connection nodes to provide computing services closer to the user, reducing latency and improving user experience. Edge computing nodes can function as general computing nodes or computing center nodes.

[0117] Edge computing connection nodes: These are nodes with connection and management functions, used as connection nodes between computing nodes, and typically not as computing nodes themselves. For example, an edge computing connection node could be a User Plane Function (UPF) node in the core network.

[0118] Core nodes: Computing nodes with centralized and powerful computing capabilities, typically corresponding to core network device nodes. Using the application of various computing resource nodes in an online training scenario as an example, when the training task is within the core network, core nodes can function as data generation / collection nodes, general computing nodes, or computing center nodes. The mobile communication system's policy library can be located on a pre-defined core node, used to store all computing power allocation policies for the mobile communication system.

[0119] The aforementioned terminals can be devices that provide voice and / or data connectivity to users, handheld devices with wireless connectivity, or other processing devices connected to a wireless modem. The name of the terminal may differ in different systems; for example, in a 5G system, the terminal can be called User Equipment (UE). Wireless terminals can communicate with one or more core networks (CNs) via a Radio Access Network (RAN). Wireless terminals can be mobile terminal devices, such as mobile phones (or "cellular" phones) and computers with mobile terminal devices, for example, portable, pocket-sized, handheld, computer-embedded, or vehicle-mounted mobile devices that exchange voice and / or data with the RAN. Examples include Personal Communication Service (PCS) phones, cordless phones, Session Initiated Protocol (SIP) phones, Wireless Local Loop (WLL) stations, and Personal Digital Assistants (PDAs). The wireless terminal can also be referred to as a system, subscriber unit, subscriber station, mobile station, mobile station, remote station, access point, remote terminal, access terminal, user terminal, user agent, or user device, but is not limited to these terms in the embodiments of this application.

[0120] The aforementioned base station may include multiple cells providing services to terminals. Depending on the specific application, a base station may also be called an access point, or a device in the access network that communicates with wireless terminals via one or more sectors on the air interface, or other names. The base station can be used to exchange received air frames with Internet Protocol (IP) packets, acting as a router between the wireless terminal and the rest of the access network, which may include an IP communication network. The base station can also coordinate the attribute management of the air interface. For example, the base station involved in the embodiments of this application can be a base transceiver station (BTS) in Global System for Mobile Communications (GSM) or Code Division Multiple Access (CDMA), or a base station (NodeB) in Wide-band Code Division Multiple Access (WCDMA), or an evolved Node B (eNB or e-NodeB) in a long-term evolution (LTE) system, a 5G base station (gNB) in a next-generation 5G network architecture, or a Home evolved Node B (HeNB), relay node, femto, pico, etc., and is not limited in the embodiments of this application. In some network structures, the base station may include a Centralized Unit (CU) node and a Distributed Unit (DU) node, and the Centralized Unit and Distributed Unit may also be geographically separated.

[0121] In this embodiment of the application, the computing center node can receive initial status information sent by multiple computing resource nodes.

[0122] Step 102: Based on the initial state information of multiple computing resource nodes, determine the target computing power allocation strategy corresponding to at least one target computing resource node among the multiple computing resource nodes.

[0123] In this embodiment of the application, the target computing resource node can be any one of the above-mentioned multiple computing resource nodes.

[0124] In this embodiment of the application, the computing center node can determine the target computing power allocation strategy corresponding to at least one target computing resource node among the multiple computing resource nodes based on the initial state information of multiple computing resource nodes.

[0125] Step 103: Send the corresponding target computing power allocation strategy to at least one target computing resource node. The target computing power allocation strategy includes the target computing power allocation ratio of multiple computing resource nodes, which is used by the corresponding target computing resource node to distribute the data to be processed according to the target computing power allocation ratio of each computing resource node.

[0126] In the embodiments of this application, the data to be processed can be training data, data related to the task to be processed, or any data that needs to be processed. This application does not impose any restrictions on this.

[0127] In this embodiment of the application, after determining the target computing power allocation strategy corresponding to each target computing resource node, the computing center node can send the corresponding target computing power allocation strategy to each target computing resource node. The target computing power allocation strategy includes the target computing power allocation ratio of the above-mentioned multiple computing resource nodes, which is used by the corresponding target computing resource nodes to distribute the data to be processed according to the target computing power allocation ratio of each computing resource node.

[0128] In other words, after receiving the corresponding target computing power allocation strategy, each target computing resource node can determine the first target data that matches the target computing power allocation ratio of the target computing resource node in the corresponding target computing power allocation strategy from the corresponding data to be processed, and send the remaining data in the data to be processed of the target computing resource node other than the first target data to the upper-level computing resource node according to the target computing power allocation ratio of the upper-level computing resource node of the target computing resource node among the above-mentioned multiple computing resource nodes.

[0129] For example, suppose the amount of data to be processed corresponding to the target computing resource node is 100, and suppose the target computing resource node has 3 superior computing resource nodes, namely node A, node B and node C, and the target computing power allocation ratios corresponding to node A, node B and node C are 10%, 20% and 30% respectively. Then the amount of data sent by the target computing resource node to node A is 100 * 10% = 10, the amount of data sent to node B is 100 * 20% = 20, and the amount of data sent to node C is 100 * 30% = 30. The amount of the first target data retained by the target computing resource node can be 100 * (100% - 10% - 20% - 30%) = 40.

[0130] In one possible implementation of this application, the target computing resource node can not only send data to its superior computing resource node, but also receive data sent by its subordinate computing resource nodes. Specifically, for each target computing resource node, it can also receive second target data sent by each of its corresponding subordinate computing resource nodes. This second target data is determined based on the target computing power allocation ratio for each subordinate computing resource node and the data to be processed for each subordinate computing resource node. In other words, for each subordinate computing resource node of the target computing resource node, the subordinate computing resource node can determine the second target data based on the corresponding data to be processed and the corresponding target computing power allocation ratio.

[0131] For example, if the amount of data to be processed corresponding to the lower-level computing resource node is 100, and the corresponding target computing power allocation ratio is 10%, then the amount of data for the second target data is 100 * 10% = 10.

[0132] As an application scenario, we will illustrate this with an example of using multiple computing resource nodes in an online training scenario. Each of these nodes can receive the computing power allocation strategy sent by the central computing node and distribute the corresponding data to be processed according to the strategy. Therefore, when performing online training tasks in a mobile communication system, the available computing resources within the system can be fully utilized to efficiently complete the online training task, thus improving the resource utilization rate of the mobile communication system.

[0133] The data processing method of this application embodiment obtains initial state information of multiple computing resource nodes; determines a target computing power allocation strategy corresponding to at least one target computing resource node among the multiple computing resource nodes based on the initial state information of the multiple computing resource nodes; and sends the corresponding target computing power allocation strategy to the at least one target computing resource node. The target computing power allocation strategy includes a target computing power allocation ratio for the multiple computing resource nodes, which is used by the corresponding target computing resource node to distribute the data to be processed according to the target computing power allocation ratio of the multiple computing resource nodes. Therefore, it is possible to generate a computing power allocation strategy corresponding to each computing resource node based on the state of each computing resource node, so that each computing resource node distributes the corresponding data to be processed according to the corresponding computing power allocation strategy. This allows for dynamic adjustment of the amount of data to be processed by each computing resource node, ensuring that the amount of data processed by each computing resource node matches its own computing power, thereby improving the resource utilization of each computing resource node in the mobile communication system and improving the data processing efficiency of the mobile communication system.

[0134] As an example, in a mobile communication system, the application of computing resource nodes in an online training scenario is used as an illustration. The interaction between the nodes can be as follows: Figure 2 As shown. Without loss of generality, Figure 2 The example only uses two nodes of the same type.

[0135] in, Figure 2 The functions of each node are as follows:

[0136] 1. End nodes are used to generate or collect raw data. Optionally, end nodes can perform preprocessing processes such as feature engineering or data augmentation on the raw data to generate training data.

[0137] End nodes can also take into account storage space, computing power, and system status to determine whether to offload training data and tasks from this node to general computing nodes (excluding edge computing connection nodes);

[0138] Execute online training computation tasks on this node;

[0139] The node status information is reported to the preset edge computing node, which serves as the computing center node, via the aggregation node and the edge computing connection node.

[0140] Receive the learned computing power allocation strategy.

[0141] 2. Aggregation nodes can serve as data generation / collection nodes, intermediate nodes, or general computing nodes.

[0142] When the aggregation node acts as a data generation / collection node, its function is as follows:

[0143] 1) Used to generate or collect raw data. Optionally, the aggregation node can perform preprocessing procedures such as feature engineering or data augmentation on the raw data to generate training data.

[0144] 2) Considering storage space, computing power, and system status, determine whether to offload the training data and tasks of this node to a general computing node (excluding edge computing connection nodes);

[0145] 3) Perform online training computation tasks at this node;

[0146] 4) The node reports its status information to the preset edge computing node, which serves as the computing center node, via the edge computing connection node;

[0147] 5) Receive the learned computing power allocation strategy;

[0148] When a pool node is used as a general computing node, its function is as follows:

[0149] 1) Training data for aggregation endpoint splitting;

[0150] 2) Perform online training computation tasks at this node;

[0151] 3) The node reports its status information to the preset edge computing node, which serves as the computing center node, via the edge computing connection node;

[0152] 4) Receive the learned computing power allocation strategy;

[0153] When the aggregation node acts as an intermediate node, its function is to forward end node status information to the preset edge computing node, which acts as the computing center node, via the edge computing connection node.

[0154] 3. Edge computing nodes can be used as general computing nodes or computing center nodes.

[0155] When an edge computing node is used as a general computing node, its function is as follows:

[0156] 1) Training data for splitting the flow between aggregation endpoints and sink nodes;

[0157] 2) Perform online training computation tasks at this node;

[0158] 3) Report the status information of this node to the preset edge computing nodes that serve as the computing center node;

[0159] 4) Receive the learned computing power allocation strategy;

[0160] When an edge computing node acts as a computing center node, its function is as follows:

[0161] 1) Driving the training process of computing power allocation strategy;

[0162] 2) Simulate the actions of data generation / collection nodes according to a certain strategy;

[0163] 3) Receive status information from data generation / acquisition nodes and general computing nodes;

[0164] 4) Execute reinforcement learning algorithms;

[0165] 5) Train and generate computing power allocation strategies;

[0166] 6) Return the computing power allocation strategy to the data generation / collection nodes and general computing nodes;

[0167] 7) Optionally, the computing power allocation strategy is sent to a preset core node that serves as the strategy library for storage.

[0168] 4. Edge computing connection nodes are nodes with connection and management functions, used as connection nodes between computing nodes, and usually not as computing nodes themselves. For example, an edge computing connection node can be a UPF functional node in the core network.

[0169] Edge computing connection nodes can be used to connect aggregation nodes, edge computing nodes, and core nodes;

[0170] Edge computing connection nodes can be used for data transmission and management between aggregation nodes, edge computing nodes, and core nodes;

[0171] Edge computing connection nodes can be used to forward training data and computing power allocation strategies between aggregation nodes, edge computing nodes, and core nodes.

[0172] 5. Core nodes can function as data generation / acquisition nodes, general computing nodes, or computing center nodes. When a core node functions as a data generation / acquisition node, its role is as follows:

[0173] 1) Generate or collect raw data;

[0174] 2) Optionally, the core node can perform preprocessing procedures such as feature engineering or data augmentation on the above-mentioned raw data to generate training data;

[0175] 3) Considering storage space, computing power, and system status, determine whether to offload the training data and tasks of this node to other core nodes;

[0176] 4) Perform online training computation tasks at this node;

[0177] 5) Report the status information of this node to the preset core node that serves as the computing center node;

[0178] 6) Receive the learned computing power allocation strategy;

[0179] When a core node is used as a general computing node, its function is as follows:

[0180] 1) Aggregate training data from other core nodes;

[0181] 2) Perform online training computation tasks at this node;

[0182] 3) Report the status information of this node to the preset core node that serves as the computing center node;

[0183] 4) Receive the learned computing power allocation strategy;

[0184] When a core node serves as the central computing node, its function is as follows:

[0185] 1) Driving the training process of computing power allocation strategy;

[0186] 2) Simulate the actions of data generation / collection nodes according to a certain strategy;

[0187] 3) Receive status information from other core nodes;

[0188] 4) Execute reinforcement learning algorithms;

[0189] 5) Train and generate computing power allocation strategies;

[0190] 6) Return the computing power allocation strategy to other core nodes;

[0191] 7) Optionally, the computing power allocation strategy is sent to a preset core node that serves as the strategy library for storage;

[0192] When the core node serves as the system's policy library, its role is as follows:

[0193] 1) Receive and save the learned computing power allocation strategy;

[0194] 2) Respond to and return the computing power allocation strategy requested.

[0195] also, Figure 2 The functions of the communication interfaces between the nodes are as follows:

[0196] Interface A: The interface between the core node and the aggregation node, used to transmit user data and control information. The user data includes computing power allocation strategy information, etc.

[0197] b interface: The interface between the aggregation node and the end nodes is used to transmit user data and control information. The user data includes training data, end node status information, and computing power allocation strategy information, etc. The b interface may be wireless or wired depending on the network conditions.

[0198] c interface: The interface between the aggregation node and the edge computing connection node is used to transmit user data, including end node and aggregation node status information, training data, and computing power allocation strategy information, etc.

[0199] d interface: The interface between aggregation nodes can be composed of common switching networks or point-to-point connections. It is used to transmit end node and aggregation node status information, training data, and computing power allocation strategy information, etc.

[0200] e-interface: The interface between edge computing nodes and edge computing connection nodes, used to transmit user data, including end node and aggregation node status information, training data, and computing power allocation strategy information, etc.

[0201] f-interface: The interface between edge computing nodes can be composed of common switching networks or point-to-point connections. It is used to transmit user data, including end node and aggregation node status information, training data, and computing power allocation strategy information, etc.

[0202] g interface: The interface between core nodes and edge computing connection nodes, used to transmit management-related data and user data, including computing power allocation strategy information.

[0203] Figure 3 This is a flowchart illustrating another data processing method provided in an embodiment of this application.

[0204] like Figure 3 As shown, this data processing method can be executed by a computing center node and may include the following steps:

[0205] Step 301: Obtain the initial status information of multiple computing resource nodes.

[0206] Step 301 can be implemented in any of the embodiments of this application. The embodiments of this application do not limit this and will not be described in detail.

[0207] In any embodiment of this application, the initial state information of each computing resource node may include at least one of the following: the amount of data to be processed by the computing resource node, the amount of computation required by the computing resource node to process the data to be processed, the available storage space of the computing resource node, and the available computing power of the computing resource node, wherein the available computing power is used to characterize the amount of data that the computing resource node can process per unit time.

[0208] As an example, assuming the number of computing resource nodes is n, the initial state information of the j-th computing resource node (hereinafter referred to as node j) is as follows:

[0209] s j,t =(d j,t ,c j,t ,m j ,g j (1)

[0210] Where 1≤j≤n, d j,t c represents the amount of data to be processed in node j, which can be in megabytes. j,t The computational cost required to process the data at node j, expressed in megabytes (Mbytes); m j The available storage space for node j, in megabytes; g j Let be the available computing power of node j, i.e., the amount of data that can be processed per unit of time (e.g., per second). Available computing power can include logical operation capability (measured in OPS, i.e., operations per second), parallel computing capability (measured in FLOPS, i.e., floating-point operations per second), and neural network acceleration capability (measured in FLOPS). A specific expression can be:

[0211]

[0212] The first term represents n in node j. j,1 The computing power of a logic operation chip, when node j contains a logic operation chip, ρ j,1 =1, otherwise, ρ j,1 =0; the second term represents n in node j. j,2 The computing power of a parallel computing chip, when node j contains a parallel computing chip, ρ j,2 =1, otherwise, ρ j,2 =0; the third term represents n in node j. j,3 The computing power of a neural network acceleration chip, when node j contains a neural network acceleration chip, ρ j,3 =1, otherwise, ρ j,3 =0; F(x) is the mapping function of chip computing power. q represents the mapping ratio, and q represents the computing power of other computing units.

[0213] Step 302: Generate a state sequence based on the initial state information of multiple computing resource nodes.

[0214] In this embodiment of the application, the computing center node can generate a state sequence based on the initial state information of multiple computing resource nodes.

[0215] As an example, the labeled state sequence is s t , then s t =(s 1,t ,s 2,t ,…,s n,t ); where n is the number of computing resource nodes.

[0216] Step 303: Generate an action sequence based on the initial state information of multiple computing resource nodes and the initial computing power allocation strategy of at least one target computing resource node; wherein, the initial computing power allocation strategy includes the initial computing power allocation ratio of multiple computing resource nodes.

[0217] In this embodiment, the computing center node can generate an action sequence based on the initial state information of multiple computing resource nodes and the initial computing power allocation strategy of at least one target computing resource node. The initial computing power allocation strategy of the target computing resource node includes the initial computing power allocation ratio of the multiple computing resource nodes. For example, the initial computing power allocation ratio of the multiple computing resource nodes in the initial computing power allocation strategy can be (1,0,...,0), or it can be a randomly assigned ratio; this application does not impose any limitations on this.

[0218] As an example, let the action sequence be labeled a t ,but:

[0219] at =(a 1,t ,…,a j,t ,…,a n,t (3)

[0220] Among them, a j,t This represents the action of node j. The amount of data d to be processed for node j is... j,t In terms of a j,t =1+∑ i∈Θ (p i→j,t ×d i,t / d j,t )-∑ k∈K p j→k,t The computational cost c required to process the data at node j. j,t In terms of a j,t =1+∑ i∈Θ (p i→j,t ×c i,t / c j,t )-∑ k∈K p j→k,t .

[0221] Where, 0≤p i→j,t ≤1 indicates the proportion of data sent from node j's lower-level node i to node j in the initial computing power allocation strategy (denoted as the initial computing power allocation proportion in this application), 0≤p j→k,t ≤1 represents the proportion of data sent by node j to its parent node k (referred to as the initial computing power allocation proportion in this application), Θ is the set of lower-level nodes that send data to node j, and K is the set of parent nodes to which node j sends data.

[0222] Step 304: Determine the cumulative reward based on the state sequence and action sequence; wherein the cumulative reward is inversely related to the amount of resources or time required to execute the action sequence in the state indicated by the state sequence.

[0223] In this embodiment of the application, the computing center node can determine the cumulative reward based on the state sequence and the action sequence, wherein the cumulative reward is inversely related to the amount of resources or time required to execute the action sequence in the state indicated by the state sequence.

[0224] As an example, using a cumulative reward as a state-action value function (such as the Q-function) as an illustration, the reward function r can be calculated based on a sequence of states and a sequence of actions. t : or,

[0225] Among them, T j,t,1 The time required to process the data to be processed for node j. In the formula G(dj,t ) is d j,t Number of computations required for the amount of data; T j,t,2 Transmit p to node j j→k,t d j,t The time required for data processing In the formula R j T represents the transmission rate of node j; j,t,3 Receive (∑) for node j i∈Θ p i→j,t d i,t The time required for data processing In the formula R i Let be the transmission rate of node i; μ is a constant greater than 0.

[0226] E j,t,1 The amount of resources (or energy) consumed in processing the data to be processed for node j. In the formula η j E represents the computing power efficiency of node j, measured in OPS / W (Operations / J). j,t,2 Transmit p to node j j→k d j,t-1 The amount of resources (or energy) consumed. In the formula ζ j E represents the energy efficiency of node j, measured in bytes / J. j,t,3 Receive (∑) for node j i∈Θ (p i→j ×d i,t-1 / d j,t-1 ))d j,t-1 The amount of resources (or energy) consumed by the data. In the formula ζ i Let i be the energy efficiency of node i, and the unit can be bytes / J.

[0227] It should be noted that the computing efficiency of node j is related to its hardware capabilities and can be measured in a certain way. For example, the computing efficiency of each node can be measured in advance using specific measurement programs or tools. The energy efficiency of node i is related to its hardware and system design and can be measured in advance. For example, the energy efficiency of node i can be estimated in advance based on the amount of test data sent by node i multiple times, its transmission power, power consumption, etc.

[0228] Therefore, in this application, the cumulative reward can be determined based on the reward function, wherein the reward function is inversely related to the amount of resources or time required to execute the action sequence in the state indicated by the state sequence, and the cumulative reward is positively related to the reward function.

[0229] As an example, let's illustrate this by using a cumulative reward as a state-action value function (such as the Q function). The Q function can be:

[0230] For ease of representation, the iteration subscript t is omitted in the above formula. Let γ be the probability of transitioning from state s to state s' by performing action a in state s, where S is a finite state space, A is a finite action space, and γ is a discount factor.

[0231] Step 305: Based on the cumulative reward, update the initial computing power allocation strategy of at least one target computing resource node at least once, and train a reinforcement learning model to obtain the target computing power allocation strategy corresponding to at least one target computing resource node.

[0232] In this embodiment of the application, the initial computing power allocation strategy of at least one target computing resource node can be updated at least once according to the cumulative reward to train the reinforcement learning model, so as to obtain the target computing power allocation strategy corresponding to at least one target computing resource node.

[0233] Step 306: Send the corresponding target computing power allocation strategy to at least one target computing resource node. The target computing power allocation strategy includes the target computing power allocation ratio of multiple computing resource nodes, which is used by the corresponding target computing resource node to distribute the data to be processed according to the target computing power allocation ratio of each computing resource node.

[0234] Step 306 can be implemented in any of the embodiments of this application. The embodiments of this application do not limit this and will not be described in detail.

[0235] The data processing method of this application embodiment generates a state sequence based on the initial state information of multiple computing resource nodes; generates an action sequence based on the initial state information of multiple computing resource nodes and the initial computing power allocation strategy of at least one target computing resource node; wherein the initial computing power allocation strategy includes the initial computing power allocation ratio of multiple computing resource nodes; determines a cumulative reward based on the state sequence and the action sequence; wherein the cumulative reward is inversely related to the amount of resources or time required to execute the action sequence in the state indicated by the state sequence; and updates the initial computing power allocation strategy of at least one target computing resource node at least once based on the cumulative reward, and trains a reinforcement learning model to obtain the target computing power allocation strategy corresponding to at least one target computing resource node. Therefore, it is possible to effectively determine the target computing power allocation strategy corresponding to a target computing resource node based on a reinforcement learning algorithm.

[0236] The following is combined with Figure 4 The training process of the reinforcement learning model is explained.

[0237] Figure 4 This is a flowchart illustrating another data processing method provided in an embodiment of this application. This data processing method can be executed by a computing center node.

[0238] like Figure 4 As shown, the reinforcement learning model is trained through the following steps:

[0239] Step 401: Obtain training samples. The training samples include state sequences of multiple computing resource nodes, action sequences corresponding to the maximum cumulative reward, rewards obtained after taking the action sequences corresponding to the maximum cumulative reward under the state sequences, and state sequences transitioned after taking the action sequences corresponding to the maximum cumulative reward.

[0240] The action sequence corresponding to the maximum cumulative reward is determined based on the computing power allocation ratio of the multiple computing resource nodes that maximize the cumulative reward.

[0241] The reward can be any value that the reward function can take.

[0242] Step 402: Train a reinforcement learning model based on the training samples to update the model parameters of the reinforcement learning model.

[0243] In this embodiment of the application, a reinforcement learning model can be trained based on training samples to update the model parameters of the reinforcement learning model.

[0244] Step 403: If the model training termination condition is not met, continue training the updated reinforcement learning model using the updated training samples until the model training termination condition is met, and then stop training the reinforcement learning model.

[0245] In this embodiment, it can be determined whether the model training termination condition is met. If the model training termination condition is not met, the training samples can be updated. For example, at least one of the action sequence, state sequence, and cumulative reward in the training samples can be updated. The updated training samples are then used to continue training the updated reinforcement learning model until the model training termination condition is met, at which point training of the reinforcement learning model stops. Thus, the trained reinforcement learning model can output a target computing power allocation strategy corresponding to at least one target computing resource node among multiple computing resource nodes. That is, the initial state information of multiple computing resource nodes can be input into the trained reinforcement learning model, and the reinforcement learning model outputs the target computing power allocation strategy corresponding to each target computing resource node.

[0246] In any embodiment of this application, the model training termination condition includes at least one of the following conditions: the cumulative reward of the reinforcement learning model converges, the training time of the reinforcement learning model reaches a preset threshold, and the number of iterations of the reinforcement learning model reaches a set number.

[0247] In any embodiment of this application, an example is given using multiple computing resource nodes in an online training scenario. The data to be processed by the multiple computing resource nodes can be used to train a target model, which can be any AI model in the related art. After the target computing resource nodes distribute the data to be processed, each computing resource node can process its own retained data and the received data to complete the training of the target model.

[0248] Specifically, after receiving the second target data sent by each of the corresponding lower-level computing resource nodes, each target computing resource node can train the target model based on the first target data it retains and the second target data sent by each of the lower-level computing resource nodes, so as to obtain the target model trained by the target computing resource node.

[0249] Other computing resource nodes can also train the target model based on the corresponding data to be processed, so as to obtain the target model after being trained by the corresponding computing resource node.

[0250] Multiple computing resource nodes can send the target model trained by the corresponding computing resource node to the computing center node. Correspondingly, the computing center node can receive the target model trained by the corresponding computing resource node sent by the multiple computing resource nodes.

[0251] The computing center node can aggregate the target models sent by the multiple computing resource nodes and trained by the corresponding computing resource nodes to obtain the aggregated target model, and then send the aggregated target model to the multiple computing resource nodes.

[0252] Alternatively, the central computing node can receive pending data from multiple computing resource nodes and train the target model based on the pending data from each computing resource node to obtain the target model trained by the central computing node. The central computing node can then aggregate the target model trained by the central computing node and the target models trained by each computing resource node to obtain an aggregated target model, and send the aggregated target model to multiple computing resource nodes.

[0253] In other words, in this application, the computing center node can not only serve as an auxiliary node for computing power allocation strategy calculation, but also as a model aggregation node in a distributed training scenario. In this application, the computing center node can aggregate the target model trained by each node and send the aggregated target model to each node, so that each node can use the aggregated target model after receiving it.

[0254] For example, taking an autoencoder as an example, the target model can include an encoder and a decoder (or decoder). During model training, each computing resource node can input the raw data from the data to be processed into the encoder, which encodes the raw data to obtain encoded data. The decoder then decodes the encoded data to obtain decoded data. The difference between the decoded data and the raw data can be calculated, and the reconstruction error can be determined based on this difference. The goal of training is to minimize the reconstruction error.

[0255] Among them, autoencoders can be used for channel state information (CSI) feedback and channel reconstruction in frequency-division duplex (FDD) systems.

[0256] The data processing method of this application embodiment trains a reinforcement learning model, thereby enabling the use of the trained reinforcement learning model to output the target computing power allocation strategy corresponding to each target computing resource node.

[0257] Figure 5 This is a flowchart illustrating another data processing method provided in an embodiment of this application.

[0258] like Figure 5 As shown, this data processing method can be executed by a computing center node and may include the following steps:

[0259] Step 501: Obtain at least one of the following: the location of each computing resource node, the data characteristics of the data to be processed on each computing resource node, and the data type.

[0260] In this embodiment, the computing center node can obtain the location of each computing resource node. For example, each computing resource node can send its location to the computing center node, thereby allowing the computing center node to obtain the location of each computing resource node.

[0261] In this embodiment, the computing center node can obtain the data type of the data to be processed from each computing resource node. For example, each computing resource node can send the data type of its corresponding data to be processed to the computing center node, thereby allowing the computing center node to obtain the data type of the data to be processed from each computing resource node.

[0262] In this embodiment of the application, the computing center node can obtain the data features of the data to be processed by each computing resource node. For example, each computing resource node can send the data features of its corresponding data to be processed to the computing center node, so that the computing center node can obtain the data features of the data to be processed by each computing resource node.

[0263] Step 502: Based on at least one of the following: the location of each computing resource node, the data characteristics of the data to be processed on each computing resource node, and the data type, determine multiple computing resource nodes from among the computing resource nodes.

[0264] In one possible implementation of this application embodiment, the computing center node can determine multiple computing resource nodes from among the computing resource nodes based on the location of each computing resource node.

[0265] As an example, the distance between each computing resource node and the computing center node can be determined based on their respective locations. Then, multiple computing resource nodes can be selected from these distances. For instance, the computing resource nodes can be sorted from smallest to largest distance, and a predetermined number of the top-ranked nodes can be selected as the multiple computing resource nodes mentioned above.

[0266] In another possible implementation of this application embodiment, the computing center node can determine multiple computing resource nodes from among the computing resource nodes according to the data type of the data to be processed by each computing resource node.

[0267] As an example, using the unprocessed data from the aforementioned multiple computing resource nodes to train the target model, the computing resource node corresponding to the unprocessed data used for training the target model can be determined from among the computing resource nodes based on the data type of the unprocessed data from each computing resource node.

[0268] For example, assuming that the data type corresponding to the training data of the target model is type A, then the computing resource nodes whose data type to be processed is type A can be determined from each computing resource node.

[0269] In another possible implementation of this application embodiment, the computing center node can determine multiple computing resource nodes from each computing resource node based on the data characteristics of the data to be processed by each computing resource node.

[0270] As an example, using the unprocessed data from the aforementioned multiple computing resource nodes to train the target model, the computing resource node corresponding to the unprocessed data used for training the target model can be determined from among the computing resource nodes based on the data characteristics of the unprocessed data from each computing resource node.

[0271] For example, taking the target model as an autoencoder used for channel state information feedback and channel reconstruction in a frequency division duplex system, the computing resource nodes related to the channel state information can be determined from each computing resource node based on the data characteristics of the data to be processed.

[0272] It should be noted that the above example only illustrates how to determine multiple computing resource nodes from among the computing resource nodes based on one of the following: the location of each computing resource node, the data characteristics of the data to be processed by each computing resource node, and the data type. In actual applications, multiple computing resource nodes can be determined by combining multiple of the following: the location of each computing resource node, the data characteristics of the data to be processed by each computing resource node, and the data type. This application does not impose any restrictions on this.

[0273] Step 503: Send indication information to multiple computing resource nodes, wherein the indication information is used to instruct the multiple computing resource nodes to send corresponding initial status information to the computing center node.

[0274] In this embodiment, the computing center node can send indication information to the plurality of computing resource nodes, wherein the indication information is used to instruct the plurality of computing resource nodes to send corresponding initial state information to the computing center node. Accordingly, each of the plurality of computing resource nodes, after receiving the indication information, can send its own initial state information to the computing center node.

[0275] Step 504: Obtain the initial status information of multiple computing resource nodes.

[0276] Step 505: Based on the initial state information of multiple computing resource nodes, determine the target computing power allocation strategy corresponding to at least one target computing resource node among the multiple computing resource nodes.

[0277] Step 506: Send the corresponding target computing power allocation strategy to at least one target computing resource node. The target computing power allocation strategy includes the target computing power allocation ratio of multiple computing resource nodes, which is used by the corresponding target computing resource node to distribute the data to be processed according to the target computing power allocation ratio of each computing resource node.

[0278] Steps 504 to 506 can be implemented in any of the embodiments of this application. The embodiments of this application do not limit this and will not be described in detail.

[0279] The data processing method of this application embodiment obtains at least one of the following: the location of each computing resource node, the data characteristics and data type of the data to be processed by each computing resource node, through a computing center node; based on the location of each computing resource node, the data characteristics and data type of the data to be processed by each computing resource node, multiple computing resource nodes are determined from the computing resource nodes; and instruction information is sent to the multiple computing resource nodes, wherein the instruction information is used to instruct the multiple computing resource nodes to send corresponding initial state information to the computing center node. Therefore, the computing center node can determine the computing resource nodes participating in collaborative computing, which improves the flexibility of the method.

[0280] In any embodiment of this application, a computing power allocation strategy corresponding to each computing resource node can be generated based on a reinforcement learning method to solve the technical problem of mismatch between computing power and data volume in related technologies.

[0281] Taking the application of multiple computing resource nodes in an online training scenario as an example, the computing power allocation strategy for each node can be determined based on the amount of data to be processed, computing power, real-time computing load, and node status conditions. Thus, each computing resource node can distribute the corresponding data to be processed according to the corresponding computing power allocation strategy. This allows for the dynamic scheduling of data to be processed (such as online training data) from each computing resource node to the most suitable computing resource node, fully utilizing the network and the capabilities of each computing resource node to execute distributed training tasks. This reduces signaling overhead for data transmission, improves computing resource utilization, and enhances online training efficiency.

[0282] For example, computing resource nodes can be data generation / collection nodes, and each data generation / collection node can offload the online training task to other nodes. Depending on the online training task, the data generation / collection node can be an end node, a aggregation node, or a core node.

[0283] Since the computing power, storage space, and network transmission bandwidth of various nodes are all limited, when a computing task is scheduled from a data generation / collection node to other nodes for execution via the network, it also involves data storage, computing, and transmission. Therefore, when multiple data generation / collection nodes participate in the training task at the same time, it involves the reasonable allocation of data transmission, storage space, and computing tasks among different nodes. The computing power allocation decisions for each data generation / collection node are mutually influential, so a collaborative computing power allocation strategy for the system is required.

[0284] Reinforcement learning algorithms can be applied to generate system computing power allocation strategies. The process of generating computing power allocation strategies can be regarded as a Markov decision process and represented by a tuple (S, A, R, p), where S is the system state space, A is the action space, R is the reward function, and p is the state transition probability.

[0285] In this application, all data generation / collection nodes can be regarded as a decision-maker, taking a joint action a at each decision time t. t =(a1,…,a n And receive a reward. System status s t It will transition to the next state s based on the transition probability. t+1 .

[0286] Each data generation / collection node makes task offloading decisions based on the current status of each computing resource node, aiming to maximize the system's cumulative reward (or cumulative benefit). The decision-making process can be initiated periodically (timed mode) or when pre-defined conditions are met (on-demand mode). The basic elements of the reinforcement learning algorithm for generating the system's computing power allocation strategy can be represented as:

[0287] 1. Intelligent Agent: This is the collection of all data generation / collection nodes, assuming there are n nodes in total. Depending on the online training task, data may be acquired from any type of node, such as end nodes, aggregation nodes, or core nodes. Therefore, any of these types of nodes could be data generation / collection nodes. During the generation of the computing power allocation strategy, the computing center node simulates the actions of the data generation / collection nodes to make decisions, but the data generation / collection nodes execute the decisions.

[0288] 2. State: The state of a node in the iteration is:

[0289] s j，t =(d j，t c j，t m j g j (1)

[0290] Where 1≤j≤n, dj,t c represents the amount of data to be trained on node j in the t-th iteration, in megabytes. j,t The computational cost required to train the task at node j in the t-th iteration is expressed in megabytes (Mbytes). j The available storage space for node j is in megabytes (Mbytes); g j Let be the available computing power of node j, i.e., the amount of information or data that can be processed per second, including logical operation capability (unit: OPS, i.e., operations per second), parallel computing capability (unit: FLOPS, i.e., floating-point operations per second), and neural network acceleration capability (unit: FLOPS). The specific expression is as follows:

[0291]

[0292] The first term represents n in node j. j,1 The computing power of a logic operation chip, when node j contains a logic operation chip, ρ j,1 =1, otherwise, ρ j,1 =0; the second term represents n in node j. j,2 The computing power of a parallel computing chip, when node j contains a parallel computing chip, ρ j,2 =1, otherwise, ρ j,2 =0; the third term represents n in node j. j,3 The computing power of a neural network acceleration chip, when node j contains a neural network acceleration chip, ρ j,3 =1, otherwise, ρ j,3 =0; F(x) is the mapping function of chip computing power. q represents the mapping ratio, and q represents the computing power of other computing units.

[0293] Node status could be the status of data generation / acquisition nodes or the status of general computing nodes. System status includes the status of all computing resource nodes, including data generation / acquisition nodes, i.e., s t =(s 1,t ,s 2,t ,…,s n,t ).

[0294] 3. Action: The joint action of all data generation / acquisition nodes in the mobile communication system at the t-th iteration is:

[0295] a t =(a 1,t ,…,a j,t ,…,a n,t (3)

[0296] Among them, a j,t This represents the action of node j. Specifically, for node j, the amount of data d to be trained...j,t In terms of a j,t =1+∑ i∈Θ (p i→j,t ×d i,t / d j,t )-∑ k∈K p j→k,t The computational cost c required for task training of node j in the t-th iteration is... j,t In terms of a j,t =1+∑ i∈Θ (p i→j,t ×c i,t / c j,t )-∑ k∈K p j→k,t Where 0≤p i→j,t ≤1 indicates the proportion of data sent from the child node i of node j to node j in the t-th iteration, 0≤p j→k,t ≤1 represents the proportion of data sent by node j to its parent node k in the t-th iteration, Θ is the set of lower-level nodes that send data to node j, and K is the set of parent nodes to which node j sends data.

[0297] It should be noted that the training data retained or distributed by the data generation / collection node at this node may be the original data generated or collected, or it may be the training data that has undergone preprocessing processes such as feature engineering or data augmentation, depending on the requirements of the training task.

[0298] 4. State Update: Node j executes action a. j,t After completing the t-th iteration, the updated state is:

[0299]

[0300] Where, p i→j,t and p j→k,t The following constraints must be met:

[0301] (1+∑ i∈Θ (p i→j,t ×d i,t / d j,t )-∑ k∈K p j→k,t )d j,t <m j ;

[0302] (1+∑ i∈Θ (p i→j,t ×c i,t / c j,t )-∑ k∈K p j→k,t )c j,t <g j ;

[0303] 5. Return function:

[0304] Among them, T j,t,1 Let be the time required for node j to complete training on the training data (i.e., the data to be processed in this application) in the t-th iteration. In the formula G(d j,t ) is d j,t Number of computations required for the amount of data; T j,t,2 For node j, transmit p in the t-th iteration j→k,t d j,t The time required for data processing In the formula R j T represents the transmission rate of node j; j,t,3 For node j to receive (∑) in the t-th iteration i∈Θ p i→j,t d i,t The time required for data processing In the formula R i Let be the transmission rate of node i; μ is a constant greater than 0;

[0305] Alternatively, a reward function:

[0306] Among them, E j,t,1 The energy consumed by node j in the t-th iteration to complete the training of the training data. In the formula η j E represents the computing power efficiency of node j, measured in OPS / W (Operations / J). j,t,2 For node j, transmit p in the t-th iteration j→k d j,t-1 The energy consumed In the formula, ζ j E represents the energy efficiency of node j, measured in bytes per liter (J). j,t,3 For node j to receive (∑) in the t-th iteration i∈Θ (p i→j ×d i,t-1 / d j,t-1 ))d j,t-1 The energy consumed by the data In the formula ζ i The energy efficiency of node i is expressed in bytes / J.

[0307] 6. State-Action Value Function Q(s,a): This refers to the cumulative reward obtained by performing action a starting from state s. Since both the state space and action space are continuous, the state-action value function Q(s,a) can be approximated using a value function. The Deep Q-Networks (DQN) algorithm with empirical replay is employed, using the function Q... φ The function Q is approximated by (s,a), where Q is the function of the function Q. φ (s,a) is usually a function with parameter φ, such as a neural network, where the output is a real number, called a Q-network.

[0308] The computing power allocation strategy learning process involves all data generation / collection nodes simultaneously. During the strategy learning cycle, each data generation / collection node reports its node status to a pre-determined computing center node, which then drives the completion of the strategy learning process.

[0309] The specific learning process for computing power allocation strategies can be as follows: Figure 6 As shown:

[0310] 1) Calculate the initial training parameters for the central node. This includes, but is not limited to, information such as the parameters of the experience pool, Q-network, and target Q-network.

[0311] 2) The computing center node receives the initial node status reports from each computing resource node.

[0312] Each computing resource node, or computing node (excluding edge computing connection nodes which are not considered computing nodes in mobile communication systems), reports its initial node state s to the computing center node. j,t This includes information such as the amount of data a node generates, the computational load required for that data, and the available storage space and computing power of the node. For data generation / collection nodes, the data amount refers to the size of the data generated or collected. For general computing nodes, the data amount refers to the size of the training data received by that node; this value is 0 in the initial state. During training, the data amount is randomly generated to speed up the training process. Optionally, the node status reporting path varies depending on the node's location.

[0313] For end nodes: via aggregation nodes and edge computing connection nodes, the data is reported to the preset edge computing node that serves as the computing center node;

[0314] For aggregation nodes: they are connected to edge computing nodes and reported to the preset edge computing node that serves as the computing center node;

[0315] For core nodes: reports are directly sent to the preset core node that serves as the computing center node.

[0316] 3) The computing center node aggregates the states of the data generation / acquisition nodes and general computing nodes into the system state s. t (Refered as a state sequence in this application).

[0317] 4) Calculate the central node under the current system state, apply the DQN algorithm with experience replay, train the Q network, and update the training parameters and the current system state until the current system state is the end state.

[0318] The computation center node employs an ε-greedy algorithm, randomly selecting the proportion of training data allocated to each data generation / collection node with a probability of ε (p). 0,t ,p 1,t ,…,p k,t This allocation ratio uses a different value for each data generation / collection node, where p 0,t p represents the proportion of training data at the data generation / collection nodes. q,t (1≤q≤k) represents the proportion of data allocated to general computing nodes for online training, and satisfies... For the current node j, p q,t The proportion p corresponding to the transmission to the parent node k j→k,t For the parent node, p q,t The proportion of data p sent from the lower-level node i to this node i→j,t With a probability of 1-ε, select the current optimal action, that is, assign each data generation / collection node a training data allocation ratio of argmax. a'∈A Q(s',a') corresponds to (p 0,t ,p 1,t ,…,p k,t The value is then calculated based on the proportion of training data allocated to each node, along with the joint action 'a'. t ;

[0319] The computing center node calculates the aggregated system state s as described above. t The return r is calculated according to formula (5) or (6). t Specifically, this includes the time or resources (such as energy) consumed in reinforcement learning training.

[0320] The computing center node updates the status of the data generation / acquisition node and the general computing node according to the above proportions and formula (4), and aggregates the status of the data generation / acquisition node and the general computing node into the system status s. t+1 ;

[0321] s t a t r t s t+1 Add to the experience pool;

[0322] Train the Q-network by sampling from the experience pool;

[0323] Update the current system status s t For s t+1 ;

[0324] Update the parameters of the Q network and the target Q network;

[0325] Determine whether the current system state is the end state. End state conditions include, but are not limited to, reaching a preset number of training iterations or a preset threshold for training time. If the system is end, proceed to step 5); otherwise, proceed to step 4.

[0326] 5) Determine whether the learning of the computing power allocation strategy should be terminated. Termination conditions include, but are not limited to, the approximation function Q. φ (s,a) indicates whether the algorithm has converged, whether the number of attempts meets the preset conditions, or whether the training time reaches the preset threshold. If the algorithm terminates, proceed to step 6); otherwise, proceed to step 2.

[0327] 6) Output computing power allocation strategy: End strategy learning and output the learned strategy. Return the computing power allocation strategy to each computing resource node.

[0328] Optionally, the computing power allocation strategy can also be sent to a preset core node, which serves as the strategy library, for storage.

[0329] The following description, in conjunction with Example A, details the use of each computing resource node for online training of the autoencoder.

[0330] Example A: Channel state information feedback and channel reconstruction based on autoencoder algorithm.

[0331] In FDD systems, the gain of Massive Multiple-Input Multiple-Output (MIMO) is obtained through CSI feedback in the downlink channel. However, in FDD systems, the amount of information transmitted back through the feedback link is limited, so it needs to be compressed in a certain way during feedback. After the base station receives the CSI feedback through the feedback link, it reconstructs the original channel information using some method. Currently, deep learning-based methods can be used for CSI feedback and channel reconstruction. For example, methods such as... Figure 7 The illustrated autoencoder algorithm compresses and reconstructs the CSI. The autoencoder includes an encoder and a decoder (or decoder) for compressing and reconstructing the CSI, respectively. At the encoder, the original channel matrix is ​​converted into a compressed representation (codeword) by learning from training data. At the decoder, the codeword is reconstructed back into the original channel matrix.

[0332] In online training of autoencoder models for CSI feedback and channel reconstruction in mobile communication systems, training data is acquired at the terminal, and the location of online training involves the issue of computing power allocation. This application utilizes a reinforcement learning method to learn a computing power allocation strategy within the mobile communication system. During online training of the CSI feedback and channel reconstruction autoencoder model, this strategy is used to schedule computing and network resources within the mobile communication system, thereby improving training efficiency. The computing nodes participating in learning the computing power allocation strategy for online training of CSI feedback and channel reconstruction can be as follows: Figure 8 As shown.

[0333] The learning process for computing power allocation strategies involves all terminals simultaneously, executing actions such as... Figure 9 The learning process is illustrated. The trigger condition for execution can be scheduled startup or on-demand startup. On-demand startup means manual startup or startup when preset conditions are met (e.g., when the amount of data to be processed by the data generation / collection node accumulates to a set threshold). For on-demand startup, the participating nodes (e.g., data generation / collection nodes) can be instructed by the computing center node sending instruction information.

[0334] like Figure 9 As shown, the learning process may specifically include:

[0335] 1) Initialize training parameters for the MEC device. This includes, but is not limited to, information such as the experience pool, Q-network, and target Q-network parameters;

[0336] 2) The terminal obtains the initial state information s j,t This includes information such as the amount of data to be collected, the computational load required for the relevant data, and the available storage space and computing power on the terminal. During the learning process, the amount of data is randomly generated to accelerate the learning process.

[0337] 3) The terminal reports its initial state information s to the base station. j,t ;

[0338] 4) The base station reports terminal status information and base station status information to the MEC device. The base station status information includes the amount of data to be trained received from the terminal (initially 0), the required computational load, the available storage space and computing power of the base station, etc.

[0339] 5) The MEC device aggregates the states of each base station and terminal into a system state s. t ;

[0340] 6) Under the current state of the terminal and the base station, the MEC device applies the DQN algorithm with experience replay to train the Q network and updates the training parameters and the current state of the terminal and the base station until the current state of the terminal and the base station is the end state.

[0341] MEC devices employ an ε-greedy algorithm, randomly selecting a proportion of training data (p) for each terminal with a probability of ε. 0,t ,p 1,t ,…,p k,t This allocation ratio uses a different value for each terminal, where p 0,t p represents the proportion of data trained on the terminal. q,t The proportion of data allocated to base stations for online training, and satisfying the following conditions. For the terminal, p q,t The proportion p corresponding to the data sent to base station k j→k,t For base stations, p q,t The proportion of data p sent from terminal i to the base station i→j,t With a probability of 1-ε, select the current optimal action, that is, allocate the training data to each terminal according to the proportion argmax. a'∈A Q(s',a') corresponds to (p 0,t ,p 1,t ,…,p k,t The value is then calculated based on the allocation ratio of training data for each terminal and base station. t ;

[0342] MEC devices are based on the above-aggregated system states s t Calculate the initial return r according to formula (5) or (6). t Specifically, this includes the signaling overhead required for the terminal to report training data, the time required for the terminal to train its own data, the signaling overhead required for the base station to forward the aggregated data to the MEC device, and the time required for the base station to train the aggregated data.

[0343] The MEC device updates the terminal and base station states according to the above ratio, and aggregates the terminal and base station states into system state s according to formula (4). t+1 ;

[0344] s t a t r t s t+1 Place it in the experience pool;

[0345] Train the Q-network by sampling from the experience pool;

[0346] Update the current aggregation status of terminals and base stations. t For s t+1 ;

[0347] Update the parameters of the Q network and the target Q network;

[0348] Determine whether the current terminal and base station status is in the "end" state. End state conditions include, but are not limited to, reaching a preset number of training iterations or a preset threshold for training time. If the training is complete, proceed to step 7); otherwise, proceed to step 6.

[0349] 7) Determine whether the learning of the computing power allocation strategy should be terminated. Termination conditions include, but are not limited to, the approximation function Q. φ (s,a) indicates whether the algorithm has converged, whether the number of attempts meets the preset conditions, or whether the training time reaches the preset threshold. If the algorithm terminates, proceed to step 8); otherwise, proceed to step 2).

[0350] 8) Generate computing power allocation strategy: End strategy learning and generate the learned strategy;

[0351] 9) The MEC device returns the computing power allocation strategy to the base station;

[0352] 10) The MEC sends the computing power allocation strategy to the preset core network device, which serves as the strategy library, for storage;

[0353] 11) The base station returns the computing power allocation strategy to the terminal.

[0354] In the above CSI feedback and channel reconstruction embodiments, the functions of each device are as follows:

[0355] 1. Terminal: A terminal node. In this embodiment, the terminal acts as a data generation / collection node, and the data is downlink channel state information. In addition to its usual functions, the terminal has the following new functions:

[0356] 1) Collect downlink channel state information as raw data;

[0357] 2) Optionally, perform preprocessing procedures such as feature engineering or data augmentation to generate training data;

[0358] 3) Based on the overall memory, computing, and system status, determine whether it is necessary to offload terminal training data and tasks to the base station or MEC (edge ​​computing) device;

[0359] 4) If some data is retained for local training, the relevant online training computation tasks are executed on the terminal;

[0360] 5) The terminal reports its status information to the preset MEC device, which serves as the computing center node, via the base station and UPF.

[0361] 6) Receive the learned computing power allocation strategy;

[0362] 2. Base Station: A type of aggregation node. In this embodiment, the base station functions as a general computing node and intermediate node. In addition to its standard functions, the base station has the following new features:

[0363] 1) Aggregate training data from terminal traffic splitting;

[0364] 2) Perform online training computation tasks at this node;

[0365] 3) The node status information is reported to the preset MEC device, which serves as the computing center node, via the UPF;

[0366] 4) The terminal status information is forwarded to the preset MEC device, which serves as the computing center node, via the UPF;

[0367] 5) Receive the learned computing power allocation strategy;

[0368] 3. MEC Device: Belonging to edge computing nodes, this is an MEC host or App server located close to the radio access network side. It connects to the base station and core network via a UPF (User-Defined Frame) and is used to provide computing services closer to the user, thereby reducing latency and improving user experience. In this embodiment, the MEC device, as a computing center node, has the following additional functions on top of its usual functionality:

[0369] 1) Driving the training process of computing power allocation strategy;

[0370] 2) Simulate terminal actions according to a certain strategy;

[0371] 3) Receive status information from the terminal and base station;

[0372] 4) Execute reinforcement learning algorithms;

[0373] 5) Train and generate computing power allocation strategies;

[0374] 6) Return the computing power allocation strategy to the base station and the terminal;

[0375] 7) Optionally, the computing power allocation strategy is sent to a preset core network device that serves as the strategy library for storage;

[0376] 4. UPF: A UPF is an edge computing connection node, serving as a connection node between base stations, MEC devices, and the core network; it is not a computing node itself. In addition to its standard functions, UPF offers the following new features:

[0377] 1) Connects base stations, MEC equipment, and the core network;

[0378] 2) Data transmission and management between base stations, MEC equipment, and the core network;

[0379] 3) Forwarding training data and computing power allocation strategies among base stations, MEC devices, and the core network;

[0380] 5. Core Network: A computing node with centralized and powerful computing capabilities. In this embodiment, it serves as the system policy library. In addition to its standard functions, the core network equipment has the following new features:

[0381] 1) Receive and save the learned computing power allocation strategy sent by the MEC device;

[0382] 2) Respond to and return the computing power allocation strategy requested.

[0383] The above Figure 8 In the mobile communication system shown, the functions of each interface are as follows:

[0384] Interface a: The interface between the core network and the base station, used to transmit user data and control information. In this application, the following interactive information is added to the interface between the core network and the base station: computing power allocation strategy information;

[0385] b-interface: The interface between the base station and the terminal, used to transmit air interface data between the base station and the terminal. In this application, the following interactive information is added to the air interface: terminal status information, training data, and computing power allocation strategy information;

[0386] c-interface: The interface between the base station and the UPF, used for transmitting user data. In this application, the following interactive information is added to the interface between the base station and the UPF: terminal and base station status information, training data, and computing power allocation strategy information.

[0387] d-interface: The interface between base stations can be constructed from a common switching network or be a point-to-point connection. In this application, the following interactive information is added to the interface between base stations: terminal and base station status information, training data, and computing power allocation strategy information.

[0388] e-interface: The interface between the MEC device and the UPF, used for transmitting user data. In this application, the following interactive information is added to the interface between the MEC device and the UPF: terminal and base station status information, training data, and computing power allocation strategy information.

[0389] f-interface: The interface between MEC devices can be constructed from a common switching network or a point-to-point connection. In this application, the following interactive information is added to the interface between MEC devices: terminal and base station status information, training data, and computing power allocation strategy information.

[0390] g-interface: Interface between core network equipment and UPF. In this application, the following interactive information has been added to the interface between the core network equipment and UPF: computing power allocation strategy information.

[0391] Therefore, when performing online training tasks in a mobile communication system, all available computing resources in the mobile communication system can be fully utilized to efficiently complete the online training task and improve the resource utilization rate of the mobile communication system.

[0392] The above-described method embodiments are executed by the computing center node. This application also proposes a data processing method executed by the target computing resource node.

[0393] Figure 10 This is a flowchart illustrating another data processing method provided in an embodiment of this application.

[0394] like Figure 10 As shown, this data processing method can be executed by the target computing resource node and may include the following steps:

[0395] Step 1001: Send the initial status information of the target computing resource node to the computing center node.

[0396] In this embodiment of the application, the target computing resource node can send corresponding initial status information to the computing center node.

[0397] In one possible implementation of this application embodiment, the target computing resource node can send corresponding initial status information to the computing center node upon receiving the indication information, wherein the indication information is used to instruct the corresponding initial status information to be sent to the computing center node.

[0398] The computing center node can determine the computing resource node to which it will send instruction information based on at least one of the following: the location of each computing resource node, the data characteristics of the data to be processed by each computing resource node, and the data type.

[0399] Step 1002: Receive the target computing power allocation strategy sent by the computing center node, wherein the target computing power allocation strategy includes the target computing power allocation ratio of multiple computing resource nodes.

[0400] In this embodiment, after receiving the initial state information of the target computing resource nodes, the computing center node can determine the target computing power allocation strategy corresponding to at least one target computing resource node among the multiple computing resource nodes based on the initial state information of the multiple computing resource nodes it has received, and send the corresponding target computing power allocation strategy to at least one target computing resource node. Correspondingly, the target computing resource node can receive the target computing power allocation strategy sent by the computing center node.

[0401] It should be noted that the explanation of the target computing power allocation strategy in any of the foregoing embodiments also applies to this embodiment. That is, the explanation of the data processing method and the description of the technical details performed by the computing center node in any of the foregoing embodiments also apply to the target computing resource node. The implementation principle is similar, and will not be repeated here.

[0402] Step 1003: Distribute the data to be processed to the target computing resource nodes according to the target computing power allocation ratio of multiple computing resource nodes.

[0403] In this embodiment of the application, after receiving the target computing power allocation strategy, the target computing resource node can distribute the data to be processed by the target computing resource node according to the target computing power allocation ratio of the above-mentioned multiple computing resource nodes in the target computing power allocation strategy.

[0404] In one possible implementation of this application embodiment, after receiving the target computing power allocation strategy, the target computing resource node can determine the first target data that matches the target computing power allocation ratio of the target computing resource node from the data to be processed of the target computing resource node, and send the remaining data in the data to be processed of the target computing resource node other than the first target data to the upper-level computing resource node according to the target computing power allocation ratio corresponding to the upper-level computing resource node of the target computing resource node among the plurality of computing resource nodes.

[0405] In one possible implementation of this application, the target computing resource node can not only send data to its superior computing resource node, but also receive data sent by its subordinate computing resource nodes. Specifically, in this application, the target computing resource node can also receive second target data sent by each of its subordinate computing resource nodes. This second target data is determined based on the target computing power allocation ratio corresponding to each subordinate computing resource node and the data to be processed corresponding to each subordinate computing resource node. In other words, for each subordinate computing resource node of the target computing resource node, the subordinate computing resource node can determine the second target data based on the corresponding data to be processed and the corresponding target computing power allocation ratio.

[0406] For example, if the amount of data to be processed corresponding to the lower-level computing resource node is 100, and the corresponding target computing power allocation ratio is 10%, then the amount of data for the second target data is 100 * 10% = 10.

[0407] In any embodiment of this application, an example is given using multiple computing resource nodes in an online training scenario. The data to be processed by the multiple computing resource nodes can be used to train a target model, which can be any AI model in the related art. After the target computing resource nodes distribute the data to be processed, each computing resource node can process its own retained data and the received data to complete the training of the target model.

[0408] Specifically, after receiving the second target data sent by each of the corresponding lower-level computing resource nodes, the target computing resource node can train the target model based on the first target data it retains and the second target data sent by each of the lower-level computing resource nodes, so as to obtain the target model trained by the target computing resource node.

[0409] The target computing resource node can also send the target model trained by the target computing resource node to the computing center node. Correspondingly, after receiving the target model trained by the corresponding computing resource node from each computing resource node, the computing center node can aggregate the target models trained by the corresponding computing resource nodes to obtain the aggregated target model, and then send the aggregated target model to each computing resource node.

[0410] Alternatively, the central computing node can receive pending data from multiple computing resource nodes and train the target model based on the pending data from each computing resource node to obtain the target model trained by the central computing node. The central computing node can then aggregate the target model trained by the central computing node and the target models trained by each computing resource node to obtain an aggregated target model, and send the aggregated target model to the aforementioned multiple computing resource nodes.

[0411] Therefore, in this application, the target computing resource node can receive the aggregated target model sent by the computing center node and use the aggregated target model.

[0412] The data processing method of this application embodiment involves sending initial state information of target computing resource nodes to a computing center node; receiving a target computing power allocation strategy sent by the computing center node, wherein the target computing power allocation strategy includes target computing power allocation ratios for multiple computing resource nodes; and distributing the data to be processed for the target computing resource nodes according to the target computing power allocation ratios for multiple computing resource nodes. This allows for the generation of a computing power allocation strategy corresponding to each computing resource node based on its state, enabling each computing resource node to distribute its corresponding data to be processed according to the corresponding computing power allocation strategy. This dynamically adjusts the amount of data to be processed by each computing resource node, ensuring that the amount of data processed by each computing resource node matches its own computing power, thereby improving the resource utilization of each computing resource node in the mobile communication system and enhancing the data processing efficiency of the mobile communication system.

[0413] The technical solutions provided in this application can be applied to various systems, especially 5G systems. For example, applicable systems include Global System for Mobile Communication (GSM), Code Division Multiple Access (CDMA), Wideband Code Division Multiple Access (WCDMA) General Packet Radio Service (GPRS), Long Term Evolution (LTE), LTE Frequency Division Duplex (FDD), LTE Time Division Duplex (TDD), Long Term Evolution Advanced (LTE-A), Universal Mobile Telecommunication System (UMTS), Worldwide Interoperability for Microwave Access (WiMAX), and 5G New Radio (NR). All of these systems include terminals and network equipment. The system may also include a core network component, such as the Evolved Packet System (EPS) or the 5G system (5GS).

[0414] To implement the above embodiments, this application also proposes a computing center node.

[0415] Figure 11 This is a schematic diagram of the structure of a computing center node provided in an embodiment of this application.

[0416] like Figure 11 As shown, the computing center node may include: transceiver 1100, processor 1110, and memory 1120.

[0417] The system includes a memory 1120 for storing computer programs; a transceiver 1100 for sending and receiving data under the control of a processor 1110; and a processor 1110 for reading the computer program from the memory 1120 and performing the following operations: acquiring initial state information of multiple computing resource nodes; determining a target computing power allocation strategy corresponding to at least one target computing resource node based on the initial state information of the multiple computing resource nodes; and sending the corresponding target computing power allocation strategy to at least one target computing resource node, wherein the target computing power allocation strategy includes a target computing power allocation ratio for the multiple computing resource nodes, and the corresponding target computing resource node distributes the data to be processed according to the target computing power allocation ratio of the multiple computing resource nodes.

[0418] Transceiver 1100 is used to receive and send data under the control of processor 1110.

[0419] Among them, Figure 11 In this context, the bus architecture can include any number of interconnected buses and bridges, specifically linking various circuits together, represented by one or more processors (processor 1110) and memory (memory 1120). The bus architecture can also link together various other circuits such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and therefore will not be described further herein. The bus interface provides an interface. The transceiver 1100 can be multiple elements, including transmitters and receivers, providing units for communicating with various other devices over transmission media, including wireless channels, wired channels, optical fibers, etc. The processor 1110 is responsible for managing the bus architecture and general processing, and the memory 1120 can store data used by the processor 1110 during operation.

[0420] The processor 1110 can be a CPU, ASIC, FPGA or CPLD, and the processor 1110 can also adopt a multi-core architecture.

[0421] In one possible implementation of this application, the initial state information includes at least one of the following: the amount of data to be processed by the corresponding computing resource node, the amount of computation required by the corresponding computing resource node to process the data to be processed, the available storage space of the corresponding computing resource node, and the available computing power of the corresponding computing resource node; wherein, the available computing power is used to characterize the amount of data that the corresponding computing resource node can process per unit time.

[0422] In one possible implementation of this application, before obtaining the initial state information of multiple computing resource nodes, the method further includes:

[0423] Obtain at least one of the following: the location of each computing resource node, the data characteristics of the data to be processed on each computing resource node, and the data type.

[0424] Based on at least one of the following: the location of each computing resource node, the data characteristics of the data to be processed by each computing resource node, and the data type, multiple computing resource nodes are determined from each computing resource node.

[0425] Send indication information to multiple computing resource nodes, wherein the indication information is used to instruct the multiple computing resource nodes to send corresponding initial status information to the computing center node.

[0426] In one possible implementation of this application, determining a target computing power allocation strategy corresponding to at least one target computing resource node among the multiple computing resource nodes based on the initial state information of multiple computing resource nodes includes:

[0427] A state sequence is generated based on the initial state information of multiple computing resource nodes;

[0428] An action sequence is generated based on the initial state information of multiple computing resource nodes and the initial computing power allocation strategy of at least one target computing resource node; wherein, the initial computing power allocation strategy includes the initial computing power allocation ratio of multiple computing resource nodes.

[0429] The cumulative reward is determined based on the state sequence and the action sequence; the cumulative reward is inversely related to the amount of resources or time required to execute the action sequence in the state indicated by the state sequence.

[0430] Based on the cumulative reward, the initial computing power allocation strategy for at least one target computing resource node is updated at least once, and a reinforcement learning model is trained to obtain the target computing power allocation strategy corresponding to at least one target computing resource node.

[0431] In one possible implementation of this application, the reinforcement learning model is trained through the following steps:

[0432] Acquire training samples, which include state sequences of multiple computing resource nodes, action sequences corresponding to the maximum cumulative reward, rewards obtained after taking the action sequences corresponding to the maximum cumulative reward under the state sequences, and state sequences transitioned after taking the action sequences corresponding to the maximum cumulative reward. The action sequences corresponding to the maximum cumulative reward are determined based on the computing power allocation ratio of the multiple computing resource nodes that maximize the cumulative reward.

[0433] A reinforcement learning model is trained based on training samples to update the model parameters of the reinforcement learning model;

[0434] If the model training termination condition is not met, the updated reinforcement learning model is trained using the updated training samples until the model training termination condition is met, at which point the training of the reinforcement learning model is stopped.

[0435] In one possible implementation of this application, the model training termination condition includes at least one of the following conditions: the cumulative reward of the reinforcement learning model converges, the training time of the reinforcement learning model reaches a preset threshold, and the number of iterations of the reinforcement learning model reaches a set number.

[0436] One possible implementation of this application also includes:

[0437] Receive pending data from multiple computing resource nodes;

[0438] The target model is trained based on the data to be processed from multiple computing resource nodes to obtain the target model after training by the computing center node.

[0439] Receive the target model trained by multiple computing resource nodes.

[0440] The target model trained by the computing center node and the target models trained by each computing resource node are aggregated to obtain the aggregated target model.

[0441] Send the aggregated target model to multiple computing resource nodes.

[0442] It should be noted that the computing center node provided in this application embodiment is capable of achieving the above-mentioned... Figures 1 to 5 All method steps implemented in the method embodiment can achieve the same technical effect. Therefore, the parts that are the same as those in the method embodiment and their beneficial effects will not be described in detail here.

[0443] To implement the above embodiments, this application also proposes a target computing resource node.

[0444] Figure 12This is a schematic diagram of the structure of a target computing resource node provided in an embodiment of this application.

[0445] like Figure 12 As shown, the target computing resource node may include: transceiver 1200, processor 1210, and memory 1220.

[0446] The system includes a memory 1220 for storing computer programs; a transceiver 1200 for sending and receiving data under the control of a processor 1210; and a processor 1210 for reading the computer program from the memory 1220 and performing the following operations: sending initial status information of the target computing resource nodes to the computing center node; receiving the target computing power allocation strategy sent by the computing center node, wherein the target computing power allocation strategy includes the target computing power allocation ratio of multiple computing resource nodes; and distributing the data to be processed to the target computing resource nodes according to the target computing power allocation ratio of the multiple computing resource nodes.

[0447] Transceiver 1200 is used to receive and send data under the control of processor 1210.

[0448] Among them, Figure 12 In this context, the bus architecture may include any number of interconnected buses and bridges, specifically linking various circuits together, represented by one or more processors (processor 1210) and memory (memory 1220). The bus architecture may also link together various other circuits such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and therefore will not be described further herein. The bus interface provides an interface. The transceiver 1200 may be multiple elements, including transmitters and receivers, providing units for communicating with various other devices over transmission media, including wireless channels, wired channels, optical fibers, etc. The processor 1210 is responsible for managing the bus architecture and general processing, and the memory 1220 may store data used by the processor 1210 during operation.

[0449] The processor 1210 can be a CPU, ASIC, FPGA or CPLD, and the processor 1210 can also adopt a multi-core architecture.

[0450] In one possible implementation of this application, before sending the initial state information of the target computing resource node to the computing center node, the method further includes:

[0451] Receive indication information sent by the computing center node, wherein the indication information is used to instruct the corresponding initial status information to be sent to the computing center node.

[0452] In one possible implementation of this application, the unprocessed data of the target computing resource nodes is distributed according to the target computing power allocation ratio of multiple computing resource nodes, including:

[0453] From the unprocessed data of the target computing resource node, determine the first target data that matches the target computing power allocation ratio corresponding to the target computing resource node;

[0454] Based on the target computing power allocation ratio of the target computing resource node's parent computing resource node among multiple computing resource nodes, send the remaining data of the target computing resource node's pending data, excluding the first target data, to the parent computing resource node.

[0455] One possible implementation of this application also includes:

[0456] The system receives second target data sent by each subordinate computing resource node corresponding to the target computing resource node. The second target data is determined based on the target computing power allocation ratio of each subordinate computing resource node and the data to be processed corresponding to each subordinate computing resource node.

[0457] One possible implementation of this application also includes:

[0458] Based on the second target data and the first target data sent by each lower-level computing resource node, the target model is trained to obtain the target model trained by the target computing resource node.

[0459] Send the target model, trained by the target computing resource node, to the computing center node;

[0460] Receive the aggregated target model sent by the computing center node.

[0461] It should be noted that the target computing resource node provided in this embodiment of the invention can achieve the above-mentioned... Figure 10 All method steps implemented in the method embodiment can achieve the same technical effect. Therefore, the parts that are the same as those in the method embodiment and their beneficial effects will not be described in detail here.

[0462] With the above Figures 1 to 5 Corresponding to the data processing method provided in the embodiments, this application also provides a data processing apparatus. Since the data processing apparatus provided in the embodiments of this application is similar to the one described above… Figures 1 to 5 The data processing method provided in the embodiments corresponds to the data processing apparatus provided in the embodiments of this application, and will not be described in detail in the embodiments of this application.

[0463] Figure 13 This is a schematic diagram of the structure of a data processing device provided in an embodiment of this application.

[0464] like Figure 13 As shown, the data processing device 1300 is applied to a computing center node and may include:

[0465] The acquisition unit 1310 is used to acquire the initial state information of multiple computing resource nodes.

[0466] The determining unit 1320 is used to determine the target computing power allocation strategy corresponding to at least one target computing resource node among the multiple computing resource nodes based on the initial state information of multiple computing resource nodes.

[0467] The sending unit 1330 is used to send a corresponding target computing power allocation strategy to at least one target computing resource node. The target computing power allocation strategy includes a target computing power allocation ratio for multiple computing resource nodes, and is used by the corresponding target computing resource node to distribute the data to be processed according to the target computing power allocation ratio of the multiple computing resource nodes.

[0468] Furthermore, in one possible implementation of this application, the initial state information includes at least one of the following: the amount of data to be processed by the corresponding computing resource node, the amount of computation required by the corresponding computing resource node to process the data to be processed, the available storage space of the corresponding computing resource node, and the available computing power of the corresponding computing resource node; wherein, the available computing power is used to characterize the amount of data that the corresponding computing resource node can process per unit time.

[0469] Furthermore, in another possible implementation of this application, the acquisition unit 1310 is also used to acquire at least one of the following: the location of each computing resource node, the data characteristics of the data to be processed by each computing resource node, and the data type.

[0470] The determining unit 1320 is further configured to determine multiple computing resource nodes from each computing resource node based on at least one of the location of each computing resource node, the data characteristics of the data to be processed by each computing resource node, and the data type.

[0471] The sending unit 1330 is also used to send indication information to multiple computing resource nodes, wherein the indication information is used to instruct the multiple computing resource nodes to send corresponding initial status information to the computing center node.

[0472] Furthermore, in another possible implementation of this application, the determining unit 1320 is specifically used for: generating a state sequence based on the initial state information of multiple computing resource nodes; generating an action sequence based on the initial state information of multiple computing resource nodes and the initial computing power allocation strategy of at least one target computing resource node; wherein the initial computing power allocation strategy includes the initial computing power allocation ratio of multiple computing resource nodes; determining a cumulative reward based on the state sequence and the action sequence; wherein the cumulative reward is inversely related to the amount of resources or time required to execute the action sequence in the state indicated by the state sequence; updating the initial computing power allocation strategy of at least one target computing resource node at least once based on the cumulative reward, and training a reinforcement learning model to obtain the target computing power allocation strategy corresponding to at least one target computing resource node.

[0473] Furthermore, in yet another possible implementation of this application, the reinforcement learning model is trained through the following steps:

[0474] The acquisition unit 1310 is also used to acquire training samples, which include state sequences of multiple computing resource nodes, action sequences corresponding to the maximum cumulative reward, rewards obtained after taking the action sequences corresponding to the maximum cumulative reward under the state sequence, and state sequences transitioned after taking the action sequences corresponding to the maximum cumulative reward. The action sequences corresponding to the maximum cumulative reward are determined based on the computing power allocation ratio of the multiple computing resource nodes that maximize the cumulative reward.

[0475] The training unit is used to train the reinforcement learning model based on the training samples, so as to update the model parameters of the reinforcement learning model.

[0476] The training unit is also used to continue training the updated reinforcement learning model with updated training samples if the model training termination condition is not met, until the model training termination condition is met and the training of the reinforcement learning model is stopped.

[0477] Furthermore, in another possible implementation of this application, the model training termination condition includes at least one of the following conditions: the cumulative reward of the reinforcement learning model converges, the training time of the reinforcement learning model reaches a preset threshold, and the number of iterations of the reinforcement learning model reaches a set number.

[0478] Furthermore, in yet another possible implementation of this application, the data processing apparatus 1300 may further include:

[0479] The receiving unit is used to receive unprocessed data sent by multiple computing resource nodes.

[0480] The training unit is also used to train the target model based on the data to be processed from multiple computing resource nodes, so as to obtain the target model after training by the computing center node.

[0481] The receiving unit is also used to receive the target model trained by the computing resource nodes and sent by multiple computing resource nodes.

[0482] The aggregation unit is used to aggregate the target model trained by the computing center node and the target models trained by each computing resource node to obtain the aggregated target model.

[0483] The sending unit 1330 is also used to send the aggregated target model to multiple computing resource nodes.

[0484] It should be noted that the data processing apparatus provided in the embodiments of this application is capable of achieving the above-mentioned... Figures 1 to 5 All method steps implemented in the method embodiment can achieve the same technical effect. Therefore, the parts that are the same as those in the method embodiment and their beneficial effects will not be described in detail here.

[0485] With the above Figure 10 Corresponding to the data processing method provided in the embodiments, this application also provides a data processing apparatus. Since the data processing apparatus provided in the embodiments of this application is similar to the one described above… Figure 10 The data processing method provided in the embodiments corresponds to the data processing apparatus provided in the embodiments of this application, and will not be described in detail in the embodiments of this application.

[0486] Figure 14 This is a schematic diagram of another data processing apparatus provided in an embodiment of this application.

[0487] like Figure 14 As shown, the data processing device 1400 is applied to the target computing resource node and may include:

[0488] The sending unit 1410 is used to send the initial status information of the target computing resource node to the computing center node.

[0489] The receiving unit 1420 is used to receive the target computing power allocation strategy sent by the computing center node, wherein the target computing power allocation strategy includes the target computing power allocation ratio of multiple computing resource nodes.

[0490] The distribution unit 1430 is used to distribute the data to be processed to the target computing resource nodes according to the target computing power allocation ratio of multiple computing resource nodes.

[0491] In one possible implementation of this application, the receiving unit 1420 is further configured to receive indication information sent by the computing center node, wherein the indication information is used to instruct the corresponding initial state information to be sent to the computing center node.

[0492] In one possible implementation of this application, the distribution unit 1430 is specifically used to: determine, from the unprocessed data of the target computing resource node, a first target data matching the target computing power allocation ratio corresponding to the target computing resource node; and send the remaining data in the unprocessed data of the target computing resource node, excluding the first target data, to the upper-level computing resource node according to the target computing power allocation ratio corresponding to the upper-level computing resource node of the target computing resource node among multiple computing resource nodes.

[0493] In one possible implementation of this application, the receiving unit 1420 is further configured to receive second target data sent by each of the lower-level computing resource nodes corresponding to the target computing resource node, wherein the second target data is determined based on the target computing power allocation ratio corresponding to each lower-level computing resource node and the data to be processed corresponding to each lower-level computing resource node.

[0494] In one possible implementation of this application, the data processing apparatus 1400 may further include:

[0495] The training unit is used to train the target model based on the second target data and the first target data sent by each lower-level computing resource node, so as to obtain the target model after training by the target computing resource node.

[0496] The sending unit is used to send the target model, which has been trained by the target computing resource node, to the computing center node.

[0497] The receiving unit 1420 is also used to receive the aggregated target model sent by the computing center node.

[0498] It should be noted that the data processing apparatus provided in the embodiments of this application is capable of achieving the above-mentioned... Figure 10 All method steps implemented in the method embodiment can achieve the same technical effect. Therefore, the parts that are the same as those in the method embodiment and their beneficial effects will not be described in detail here.

[0499] It should be noted that the functional units in the various embodiments of this application 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. The integrated unit can be implemented in hardware or as a software functional unit.

[0500] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a processor-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0501] To implement the above embodiments, this application also proposes a processor-readable storage medium.

[0502] The processor-readable storage medium stores a computer program that is used to cause the processor to execute this application. Figures 1 to 5 Data processing method of any embodiment.

[0503] The processor-readable storage medium can be any available medium or data storage device that the processor can access, including but not limited to magnetic storage (e.g., floppy disks, hard disks, magnetic tapes, magneto-optical disks (MOs)), optical storage (e.g., CDs, DVDs, BDs, HVDs), and semiconductor storage (e.g., ROMs, EPROMs, EEPROMs, non-volatile memory (NAND flash), solid-state drives (SSDs)).

[0504] To implement the above embodiments, this application also proposes a processor-readable storage medium.

[0505] The processor-readable storage medium stores a computer program that is used to cause the processor to execute this application. Figure 10 The data processing method described in the embodiments.

[0506] The processor-readable storage medium can be any available medium or data storage device that the processor can access, including but not limited to magnetic memory (e.g., floppy disk, hard disk, magnetic tape, magneto-optical disk (MO)), optical memory (e.g., CD, DVD, BD, HVD), and semiconductor memory (e.g., ROM, EPROM, EEPROM, non-volatile memory (NAND FLASH), solid-state drive (SSD)).

[0507] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage and optical storage) containing computer-usable program code.

[0508] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer-executable instructions. These computer-executable instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0509] These processor-executable instructions may also be stored in a processor-readable memory that can direct a computer or other programmable data processing device to operate in a particular manner, such that the instructions stored in the processor-readable memory produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0510] These processors can execute instructions that can also be loaded onto a computer or other programmable data processing device, causing a series of operational steps to be performed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable device for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0511] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A data processing method, characterized in that, Applied to a computing center node, the method includes: Obtain the initial status information of multiple computing resource nodes; Based on the initial state information of the plurality of computing resource nodes, determine the target computing power allocation strategy corresponding to at least one target computing resource node among the plurality of computing resource nodes; Send a corresponding target computing power allocation strategy to the at least one target computing resource node, wherein the target computing power allocation strategy includes the target computing power allocation ratio of the plurality of computing resource nodes, and is used for the corresponding target computing resource node to distribute the data to be processed according to the target computing power allocation ratio of the plurality of computing resource nodes; The step of determining the target computing power allocation strategy corresponding to at least one target computing resource node among the plurality of computing resource nodes based on the initial state information of the plurality of computing resource nodes includes: A state sequence is generated based on the initial state information of the multiple computing resource nodes; An action sequence is generated based on the initial state information of the plurality of computing resource nodes and the initial computing power allocation strategy of the at least one target computing resource node; wherein, the initial computing power allocation strategy includes the initial computing power allocation ratio of the plurality of computing resource nodes; the action sequence includes the actions of the plurality of computing resource nodes; the action of node j in the action sequence is obtained based on the proportion of data sent from node i to node j, the amount of data to be processed corresponding to node i, the amount of data to be processed corresponding to node j, and the proportion of data sent from node j to node k, as specified in the initial computing power allocation strategy; or, it is obtained based on the proportion of data sent from node i to node j, the amount of computation required by node i to process the data to be processed, the amount of computation required by node j to process the data to be processed, and the proportion of data sent from node j to node k, as specified in the initial computing power allocation strategy. Based on the state sequence and the action sequence, a cumulative reward is determined; wherein the cumulative reward is inversely related to the amount of resources or time required to execute the action sequence in the state indicated by the state sequence. Based on the accumulated reward, the initial computing power allocation strategy of the at least one target computing resource node is updated at least once, and a reinforcement learning model is trained to obtain the target computing power allocation strategy corresponding to the at least one target computing resource node.

2. The method according to claim 1, characterized in that, The initial state information includes at least one of the following: the amount of data to be processed on the corresponding computing resource node, the amount of computation required by the corresponding computing resource node to process the data to be processed, the available storage space of the corresponding computing resource node, and the available computing power of the corresponding computing resource node. The available computing power is used to characterize the amount of data that the corresponding computing resource node can process per unit time.

3. The method according to claim 1, characterized in that, Before obtaining the initial state information of multiple computing resource nodes, the method further includes: Obtain at least one of the following: the location of each computing resource node, the data characteristics of the data to be processed by each computing resource node, and the data type. The plurality of computing resource nodes are determined from the computing resource nodes based on at least one of the location of each computing resource node, the data characteristics of the data to be processed by each computing resource node, and the data type. Send indication information to the plurality of computing resource nodes, wherein the indication information is used to instruct the plurality of computing resource nodes to send corresponding initial status information to the computing center node.

4. The method according to claim 1, characterized in that, The reinforcement learning model is trained through the following steps: Obtain training samples, which include the state sequence of the plurality of computing resource nodes, the action sequence corresponding to the maximum cumulative reward, the reward obtained after taking the action sequence corresponding to the maximum cumulative reward under the state sequence, and the state sequence after taking the action sequence corresponding to the maximum cumulative reward, wherein the action sequence corresponding to the maximum cumulative reward is determined according to the computing power allocation ratio of the plurality of computing resource nodes that maximizes the cumulative reward; A reinforcement learning model is trained based on the training samples to update the model parameters of the reinforcement learning model; If the model training termination condition is not met, the updated reinforcement learning model is trained using the updated training samples until the model training termination condition is met, at which point the training of the reinforcement learning model is stopped.

5. The method according to claim 4, characterized in that, The training termination conditions of the model include at least one of the following: the cumulative reward of the reinforcement learning model converges, the training time of the reinforcement learning model reaches a preset threshold, and the number of iterations of the reinforcement learning model reaches a set number.

6. The method according to any one of claims 1-5, characterized in that, The method further includes: Receive the data to be processed sent by the plurality of computing resource nodes; The target model is trained based on the data to be processed from the multiple computing resource nodes to obtain the target model after training by the computing center node. Receive the target model trained by the computing resource nodes sent by the plurality of computing resource nodes; The target model trained by the computing center node and the target models trained by each computing resource node are aggregated to obtain the aggregated target model. The aggregated target model is sent to the multiple computing resource nodes.

7. A data processing method, characterized in that, Applied to target computing resource nodes, the method includes: Send the initial status information of the target computing resource node to the computing center node; The system receives a target computing power allocation strategy sent by the computing center node, wherein the target computing power allocation strategy includes a target computing power allocation ratio for multiple computing resource nodes. The data to be processed on the target computing resource nodes is distributed according to the target computing power allocation ratio of the multiple computing resource nodes. The target computing power allocation strategy is as follows: the computing center node generates a state sequence based on the initial state information of the plurality of computing resource nodes; it generates an action sequence based on the initial state information of the plurality of computing resource nodes and the initial computing power allocation strategy of at least one target computing resource node; the initial computing power allocation strategy includes the initial computing power allocation ratio of the plurality of computing resource nodes; a cumulative reward is determined based on the state sequence and the action sequence; the cumulative reward is inversely related to the amount of resources or time required to execute the action sequence in the state indicated by the state sequence; and the initial computing power allocation strategy of the at least one target computing resource node is updated at least once based on the cumulative reward, and a reinforcement learning model is trained accordingly. The action sequence includes actions of multiple computing resource nodes. The action of node j in the action sequence is obtained based on the proportion of data sent from node i to node j in the initial computing power allocation strategy, the amount of data to be processed corresponding to node i, the amount of data to be processed corresponding to node j, and the proportion of data sent from node j to node k, the parent node k of node j; or, it is obtained based on the proportion of data sent from node i to node j in the initial computing power allocation strategy, the amount of computation required by node i to process the data to be processed, the amount of computation required by node j to process the data to be processed, and the proportion of data sent from node j to node k, the parent node k of node j.

8. The method according to claim 7, characterized in that, Before sending the initial state information of the target computing resource node to the computing center node, the method further includes: The system receives an instruction message sent by the computing center node, wherein the instruction message is used to instruct the corresponding initial state information to be sent to the computing center node.

9. The method according to claim 7 or 8, characterized in that, The step of distributing the data to be processed to the target computing resource nodes according to the target computing power allocation ratio of the multiple computing resource nodes includes: From the unprocessed data of the target computing resource node, determine the first target data that matches the target computing power allocation ratio corresponding to the target computing resource node; Based on the target computing power allocation ratio corresponding to the parent computing resource node of the target computing resource node among the plurality of computing resource nodes, the remaining data in the unprocessed data of the target computing resource node, excluding the first target data, is sent to the parent computing resource node.

10. The method according to claim 9, characterized in that, The method further includes: The system receives second target data sent by each subordinate computing resource node corresponding to the target computing resource node, wherein the second target data is determined based on the target computing power allocation ratio corresponding to each subordinate computing resource node and the data to be processed corresponding to each subordinate computing resource node.

11. The method according to claim 10, characterized in that, The method further includes: The target model is trained based on the second target data and the first target data sent by each of the lower-level computing resource nodes to obtain the target model trained by the target computing resource nodes. Send the target model, trained by the target computing resource node, to the computing center node; Receive the aggregated target model sent by the computing center node.

12. A computing center node, characterized in that, Includes memory, transceiver, and processor: A memory for storing computer programs; a transceiver for sending and receiving data under the control of the processor; and a processor for reading the computer programs from the memory and performing the following operations: Obtain the initial status information of multiple computing resource nodes; Based on the initial state information of the plurality of computing resource nodes, determine the target computing power allocation strategy corresponding to at least one target computing resource node among the plurality of computing resource nodes; Send a corresponding target computing power allocation strategy to the at least one target computing resource node, wherein the target computing power allocation strategy includes the target computing power allocation ratio of the plurality of computing resource nodes, and is used for the corresponding target computing resource node to distribute the data to be processed according to the target computing power allocation ratio of the plurality of computing resource nodes; The step of determining the target computing power allocation strategy corresponding to at least one target computing resource node among the plurality of computing resource nodes based on the initial state information of the plurality of computing resource nodes includes: A state sequence is generated based on the initial state information of the multiple computing resource nodes; An action sequence is generated based on the initial state information of the plurality of computing resource nodes and the initial computing power allocation strategy of the at least one target computing resource node; wherein, the initial computing power allocation strategy includes the initial computing power allocation ratio of the plurality of computing resource nodes; the action sequence includes the actions of the plurality of computing resource nodes; the action of node j in the action sequence is obtained based on the proportion of data sent from node i to node j, the amount of data to be processed corresponding to node i, the amount of data to be processed corresponding to node j, and the proportion of data sent from node j to node k, as specified in the initial computing power allocation strategy; or, it is obtained based on the proportion of data sent from node i to node j, the amount of computation required by node i to process the data to be processed, the amount of computation required by node j to process the data to be processed, and the proportion of data sent from node j to node k, as specified in the initial computing power allocation strategy. Based on the state sequence and the action sequence, a cumulative reward is determined; wherein the cumulative reward is inversely related to the amount of resources or time required to execute the action sequence in the state indicated by the state sequence. Based on the accumulated reward, the initial computing power allocation strategy of the at least one target computing resource node is updated at least once, and a reinforcement learning model is trained to obtain the target computing power allocation strategy corresponding to the at least one target computing resource node.

13. The computing center node according to claim 12, characterized in that, The initial state information includes at least one of the following: the amount of data to be processed on the corresponding computing resource node, the amount of computation required by the corresponding computing resource node to process the data to be processed, the available storage space of the corresponding computing resource node, and the available computing power of the corresponding computing resource node. The available computing power is used to characterize the amount of data that the corresponding computing resource node can process per unit time.

14. The computing center node according to claim 12, characterized in that, Before obtaining the initial state information of multiple computing resource nodes, the process also includes: Obtain at least one of the following: the location of each computing resource node, the data characteristics of the data to be processed by each computing resource node, and the data type. The plurality of computing resource nodes are determined from the computing resource nodes based on at least one of the location of each computing resource node, the data characteristics of the data to be processed by each computing resource node, and the data type. Send indication information to the plurality of computing resource nodes, wherein the indication information is used to instruct the plurality of computing resource nodes to send corresponding initial status information to the computing center node.

15. The computing center node according to claim 12, characterized in that, The reinforcement learning model is trained through the following steps: Obtain training samples, which include the state sequence of the plurality of computing resource nodes, the action sequence corresponding to the maximum cumulative reward, the reward obtained after taking the action sequence corresponding to the maximum cumulative reward under the state sequence, and the state sequence after taking the action sequence corresponding to the maximum cumulative reward, wherein the action sequence corresponding to the maximum cumulative reward is determined according to the computing power allocation ratio of the plurality of computing resource nodes that maximizes the cumulative reward; A reinforcement learning model is trained based on the training samples to update the model parameters of the reinforcement learning model; If the model training termination condition is not met, the updated reinforcement learning model is trained using the updated training samples until the model training termination condition is met, at which point the training of the reinforcement learning model is stopped.

16. The computing center node according to claim 15, characterized in that, The training termination conditions of the model include at least one of the following: the cumulative reward of the reinforcement learning model converges, the training time of the reinforcement learning model reaches a preset threshold, and the number of iterations of the reinforcement learning model reaches a set number.

17. The computing center node according to any one of claims 12-16, characterized in that, Also includes: Receive the data to be processed sent by the plurality of computing resource nodes; The target model is trained based on the data to be processed from the multiple computing resource nodes to obtain the target model after training by the computing center node. Receive the target model trained by the computing resource nodes sent by the plurality of computing resource nodes; The target model trained by the computing center node and the target models trained by each computing resource node are aggregated to obtain the aggregated target model. The aggregated target model is sent to the multiple computing resource nodes.

18. A target computing resource node, characterized in that, Includes memory, transceiver, and processor: A memory for storing computer programs; a transceiver for sending and receiving data under the control of the processor; and a processor for reading the computer programs from the memory and performing the following operations: Send the initial status information of the target computing resource node to the computing center node; The system receives a target computing power allocation strategy sent by the computing center node, wherein the target computing power allocation strategy includes a target computing power allocation ratio for multiple computing resource nodes. The data to be processed on the target computing resource nodes is distributed according to the target computing power allocation ratio of the multiple computing resource nodes. The target computing power allocation strategy is as follows: the computing center node generates a state sequence based on the initial state information of the plurality of computing resource nodes; it generates an action sequence based on the initial state information of the plurality of computing resource nodes and the initial computing power allocation strategy of at least one target computing resource node; the initial computing power allocation strategy includes the initial computing power allocation ratio of the plurality of computing resource nodes; a cumulative reward is determined based on the state sequence and the action sequence; the cumulative reward is inversely related to the amount of resources or time required to execute the action sequence in the state indicated by the state sequence; and the initial computing power allocation strategy of the at least one target computing resource node is updated at least once based on the cumulative reward, and a reinforcement learning model is trained accordingly. The action sequence includes actions of multiple computing resource nodes. The action of node j in the action sequence is obtained based on the proportion of data sent from node i to node j in the initial computing power allocation strategy, the amount of data to be processed corresponding to node i, the amount of data to be processed corresponding to node j, and the proportion of data sent from node j to node k, the parent node k of node j; or, it is obtained based on the proportion of data sent from node i to node j in the initial computing power allocation strategy, the amount of computation required by node i to process the data to be processed, the amount of computation required by node j to process the data to be processed, and the proportion of data sent from node j to node k, the parent node k of node j.

19. The target computing resource node according to claim 18, characterized in that, Before sending the initial state information of the target computing resource node to the computing center node, the method further includes: The system receives an instruction message sent by the computing center node, wherein the instruction message is used to instruct the corresponding initial state information to be sent to the computing center node.

20. The target computing resource node according to claim 18 or 19, characterized in that, The step of distributing the data to be processed to the target computing resource nodes according to the target computing power allocation ratio of the multiple computing resource nodes includes: From the unprocessed data of the target computing resource node, determine the first target data that matches the target computing power allocation ratio corresponding to the target computing resource node; Based on the target computing power allocation ratio corresponding to the parent computing resource node of the target computing resource node among the plurality of computing resource nodes, the remaining data in the unprocessed data of the target computing resource node, excluding the first target data, is sent to the parent computing resource node.

21. The target computing resource node according to claim 20, characterized in that, Also includes: The system receives second target data sent by each subordinate computing resource node corresponding to the target computing resource node, wherein the second target data is determined based on the target computing power allocation ratio corresponding to each subordinate computing resource node and the data to be processed corresponding to each subordinate computing resource node.

22. The target computing resource node according to claim 21, characterized in that, Also includes: The target model is trained based on the second target data and the first target data sent by each of the lower-level computing resource nodes to obtain the target model trained by the target computing resource nodes. Send the target model, trained by the target computing resource node, to the computing center node; Receive the aggregated target model sent by the computing center node.

23. A data processing apparatus, characterized in that, include: The acquisition unit is used to acquire the initial state information of multiple computing resource nodes; The determining unit is used to determine the target computing power allocation strategy corresponding to at least one target computing resource node among the plurality of computing resource nodes based on the initial state information of the plurality of computing resource nodes. A sending unit is configured to send a corresponding target computing power allocation strategy to the at least one target computing resource node, wherein the target computing power allocation strategy includes the target computing power allocation ratio of the plurality of computing resource nodes, and the corresponding target computing resource node distributes the data to be processed according to the target computing power allocation ratio of the plurality of computing resource nodes; The determining unit is specifically used for: A state sequence is generated based on the initial state information of the multiple computing resource nodes; An action sequence is generated based on the initial state information of the plurality of computing resource nodes and the initial computing power allocation strategy of the at least one target computing resource node; wherein, the initial computing power allocation strategy includes the initial computing power allocation ratio of the plurality of computing resource nodes; the action sequence includes the actions of the plurality of computing resource nodes; the action of node j in the action sequence is obtained based on the proportion of data sent from node i to node j, the amount of data to be processed corresponding to node i, the amount of data to be processed corresponding to node j, and the proportion of data sent from node j to node k, as specified in the initial computing power allocation strategy; or, it is obtained based on the proportion of data sent from node i to node j, the amount of computation required by node i to process the data to be processed, the amount of computation required by node j to process the data to be processed, and the proportion of data sent from node j to node k, as specified in the initial computing power allocation strategy. Based on the state sequence and the action sequence, a cumulative reward is determined; wherein the cumulative reward is inversely related to the amount of resources or time required to execute the action sequence in the state indicated by the state sequence. Based on the accumulated reward, the initial computing power allocation strategy of the at least one target computing resource node is updated at least once, and a reinforcement learning model is trained to obtain the target computing power allocation strategy corresponding to the at least one target computing resource node.

24. A data processing apparatus, characterized in that, include: The sending unit is used to send the initial status information of the target computing resource node to the computing center node; A receiving unit is configured to receive a target computing power allocation strategy sent by the computing center node, wherein the target computing power allocation strategy includes a target computing power allocation ratio for multiple computing resource nodes. The distribution unit is used to distribute the data to be processed by the target computing resource nodes according to the target computing power allocation ratio of the multiple computing resource nodes; The target computing power allocation strategy is as follows: the computing center node generates a state sequence based on the initial state information of the plurality of computing resource nodes; it generates an action sequence based on the initial state information of the plurality of computing resource nodes and the initial computing power allocation strategy of at least one target computing resource node; the initial computing power allocation strategy includes the initial computing power allocation ratio of the plurality of computing resource nodes; a cumulative reward is determined based on the state sequence and the action sequence; the cumulative reward is inversely related to the amount of resources or time required to execute the action sequence in the state indicated by the state sequence; and the initial computing power allocation strategy of the at least one target computing resource node is updated at least once based on the cumulative reward, and a reinforcement learning model is trained accordingly. The action sequence includes actions of multiple computing resource nodes. The action of node j in the action sequence is obtained based on the proportion of data sent from node i to node j in the initial computing power allocation strategy, the amount of data to be processed corresponding to node i, the amount of data to be processed corresponding to node j, and the proportion of data sent from node j to node k, the parent node k of node j; or, it is obtained based on the proportion of data sent from node i to node j in the initial computing power allocation strategy, the amount of computation required by node i to process the data to be processed, the amount of computation required by node j to process the data to be processed, and the proportion of data sent from node j to node k, the parent node k of node j.

25. A processor-readable storage medium, characterized in that, The processor-readable storage medium stores a computer program that causes the processor to perform the method according to any one of claims 1 to 6, or to perform the method according to any one of claims 7 to 11.

Images