A data processing system and method based on reinforcement learning

Abstract

The application relates to a data processing system and method based on reinforcement learning, which comprises a model construction module, a reinforcement learning decision module and an output module.The model construction module is used for reinforcement learning modeling for data to be processed, and is used for determining an action space and a value function, and a policy type.The reinforcement learning decision module is used for acquiring each action in the reinforcement learning model, the corresponding value, the policy type and a preset accuracy requirement, and is used for realizing rapid decision by using a hardware circuit.The output module is used for outputting a decision result, and obtaining a processing result of the data to be processed.The application realizes random policy selection by using hardware, can perform policy selection of the retention accuracy, utilization rate and exploration degree in parallel in the hardware, and improves the operation speed, so that the application can be widely applied to the field of data processing.

Description

Technical Field

[0001] This invention belongs to the field of data processing, specifically relating to the hardware acceleration of policy selection in machine learning, and particularly to a data processing system and method based on reinforcement learning. Background Technology

[0002] Reinforcement learning (RL) is a machine learning method whose core principle is to learn how to make decisions in a task through the interaction of an agent with the environment, in order to maximize the expected cumulative reward. Its key elements include:

[0003] Agent and Environment: In reinforcement learning, there is an agent, which is a learner or decision-maker, that interacts with a dynamic environment. The agent receives rewards or penalties based on its behavior in the environment, thereby learning how to make decisions in different situations.

[0004] State: A state is a way of describing an environment, containing information about the current state of that environment. A state can be discrete (e.g., the position of a piece on a chessboard) or continuous (e.g., sensor data from a robot).

[0005] Action: An action is a behavior or decision that an agent can take in a given state. The set of actions is usually discrete or continuous, depending on the specific problem.

[0006] Policy: A policy is a decision rule for an agent; it is a mapping from states to actions. A policy defines how an agent chooses actions in different states. Policies can be deterministic (each state has a definite action, generally choosing the action with the highest value) or stochastic (each state has a certain probability of choosing a different action). While deterministic policy selection is computationally efficient, it is prone to getting trapped in local optima and has low exploration space. Completely stochastic policy selection, although computationally inefficient, uses random sampling according to the probability distribution, which can increase the exploration space while preserving computational efficiency, and is considered a more promising policy selection method.

[0007] Reward: A reward is a feedback signal provided by the environment to an agent, used to evaluate the agent's behavior. The agent's goal is to maximize the cumulative reward. A reward can be positive (positive), negative (negative), or zero, representing the degree of merit of an action taken by the agent in a given state.

[0008] Value Function: A value function measures the value or expected cumulative reward of taking different actions in different states. There are two main types of value functions: State-Value Function: Represents the expected cumulative reward of taking a policy in a given state. Action-Value Function: Represents the expected cumulative reward of taking a certain action in a given state and then continuing to take actions according to the policy.

[0009] like Figure 1 The diagram illustrates the overall computational flow of reinforcement learning. Reinforcement learning algorithms, which facilitate decision-making, have broad application prospects in the field of artificial intelligence, especially in edge computing. For example, they can be used to train cars and robots to make decisions and act in complex environments to achieve autonomous navigation and task execution; they can also be used to optimize energy management, supply chain management, and environmental protection to improve resource utilization efficiency.

[0010] However, most current reinforcement learning strategy selections involve directly calling functions from libraries within software algorithms, and there is no way to implement them on hardware. Summary of the Invention

[0011] To address the issues of policy selection in the aforementioned machine learning algorithms, the purpose of this invention is to provide an efficient reinforcement learning decision-making device and its usage method, which implements a general solution for random policy selection through hardware.

[0012] To achieve the above objectives, the present invention adopts the following technical solution:

[0013] In a first aspect, the present invention provides a data processing system based on reinforcement learning, comprising:

[0014] The model building module is used to perform reinforcement learning modeling on the data to be processed, and to determine the action space, its value function, and policy type.

[0015] The reinforcement learning decision module is used to obtain each action in the reinforcement learning model and its corresponding value, policy type and preset accuracy requirements, and to realize fast decision-making using hardware circuits.

[0016] The output module is used to output the decision results and obtain the processing results of the data to be processed.

[0017] Furthermore, the strategy types in the model building module include at least: deterministic strategy, completely random strategy, ∈-greedy strategy, or probability distribution strategy.

[0018] Furthermore, the reinforcement learning decision-making module includes:

[0019] The input module is used to obtain each action in the reinforcement learning model and its corresponding value, policy type, and accuracy requirements;

[0020] The computation module is used to process the input actions and values ​​according to the strategy type and accuracy requirements to obtain the decision results.

[0021] Furthermore, the operation calculation module includes: a data processing module, a random number generation module, a sampling module, a verification module, and an initialization module;

[0022] The data processing module is used to determine and quantify the probability values ​​of each action according to the strategy type and accuracy requirements, and generate an array corresponding to the action probability values.

[0023] The random number generation module is used to randomly generate random numbers within a preset range;

[0024] The sampling module is used to sample from the array corresponding to the action probability value according to the random number, and output the action to which the tag corresponding to the random number belongs;

[0025] The verification module is used to verify the action output by the sampling module. If the preset conditions are not met, the module returns to the random number generation module to regenerate random numbers.

[0026] The initialization module is used to clear the data processing module and the verification module to 0 after the calculation is completed.

[0027] Furthermore, the data processing module includes a preprocessing module, a quantization module, and an array generation module;

[0028] The preprocessing module is used to determine the probability value corresponding to each action based on the strategy type;

[0029] The quantization module is used to quantize the probability value of each action according to the accuracy requirements, and obtain a quantization result with a preset width.

[0030] The array generation module is used to divide the probability value range into M1 intervals according to the accuracy requirements, and generate an array corresponding to the probability value of each action by comparing the interval range of the probability value corresponding to each action.

[0031] Furthermore, in the preprocessing module, when determining the probability value corresponding to each action based on the strategy type:

[0032] For deterministic strategies, the probability of the action with the highest value is set to 1, and the probabilities of all other actions are set to 0.

[0033] For a completely random strategy, the probability of each of the N actions is set to 1 / N;

[0034] For the ∈-greedy strategy, the probability of the action with the largest value is set to The probability values ​​of the other (N-1) actions are all set to...

[0035] For strategies that use the Softmax function for normalization and then make decisions based on probability distributions, the probability value corresponding to each action is set to...

[0036] Furthermore, the array generation module includes an interval division module, a comparison module, and a sorting and splicing module;

[0037] The interval division module is used to divide the probability value range into M1 intervals according to the accuracy requirements;

[0038] The comparison module is used to compare the range of probability values ​​corresponding to each action and generate N arrays consisting of 0 or 1 signals.

[0039] The sorting and splicing module is used to splice N groups of M1-bit arrays sequentially from low to high to obtain a total array of (N×M1) bits.

[0040] Furthermore, the random number generation module is connected to any hardware circuit capable of randomly generating 0 and 1 through a port to generate random numbers in the range of 0 to ((N×M1)-1).

[0041] Furthermore, the verification module includes a counter module and a judgment module. The counter module is used to count the current number of samples. The judgment module is used to determine whether the sampling was successful and whether the current number of samples has reached a preset threshold. The determination of whether the sampling was successful means: based on the generated random number, find the value corresponding to the position as 0 or 1. If it is 0, return to the random number generation module, regenerate the random number and sample again. If it is 1, output the action corresponding to the random number.

[0042] Secondly, the present invention provides a data processing method based on reinforcement learning, comprising:

[0043] Reinforcement learning modeling is performed on the data to be processed to determine the action space, its value function, and policy type.

[0044] The algorithm acquires each action in the reinforcement learning model, along with its corresponding value, policy type, and preset accuracy requirements. It then utilizes hardware circuitry to make rapid decisions and obtain the processing results for the data to be processed.

[0045] The present invention has the following advantages due to the adoption of the above technical solutions:

[0046] 1. This invention achieves hardware acceleration for the software computation part of policy selection in machine learning, enabling parallel policy selection that preserves accuracy, utilization, and exploration in hardware, thereby improving computation speed.

[0047] 2. The circuit design scheme proposed in this invention can be applied to various scenarios such as multiple strategy selection schemes and multiple inputs, and has strong versatility.

[0048] 3. Although this invention is a hardware implementation scheme specifically proposed for policy selection in machine learning, it is still applicable to other application scenarios involving probability distribution sampling.

[0049] Therefore, this invention can be widely applied in the field of data processing. Attached Figure Description

[0050] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts. In the drawings:

[0051] Figure 1 This is a background illustration of the present invention;

[0052] Figure 2 This is an overall module block diagram of the hardware implementation device of the present invention;

[0053] Figure 3 This is a schematic diagram of the internal structure of the data processing module of the present invention;

[0054] Figure 4 This is a schematic diagram of the quantization and array generation module, the random number generation module, and the interaction between the two and the sampling module of the present invention.

[0055] Figure 5 This is a schematic diagram of the internal structure of the verification module of the present invention and a schematic diagram of its relationship with the random number generation module;

[0056] Figure 6 This is a schematic diagram of a specific embodiment of the present invention;

[0057] Figure 7 This is a schematic diagram of the quantization process in a specific embodiment of the present invention. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention are within the scope of protection of the present invention.

[0059] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0060] In some embodiments of the present invention, a data processing system based on reinforcement learning is provided. The hardware implementation of several basic reinforcement learning strategy selection schemes is described. However, this type of circuit that can be used for probability distribution sampling is not limited to the field of machine learning. It can be applied to any application scenario that requires sampling according to probability distribution.

[0061] Correspondingly, in some other embodiments of the present invention, a data processing method based on reinforcement learning is also provided.

[0062] Example 1

[0063] This embodiment provides a data processing system based on reinforcement learning, which includes:

[0064] The model building module is used to perform reinforcement learning modeling on the data to be processed, and to determine the action space and its value function, policy type, etc.

[0065] The reinforcement learning decision module is used to obtain each action in the reinforcement learning model and its corresponding value, policy type and preset accuracy requirements, and to realize fast decision-making using hardware circuits.

[0066] The output module is used to output the decision results and obtain the processing results of the data to be processed.

[0067] Preferably, the strategy types determined in the model building module include, but are not limited to, the following:

[0068] Deterministic strategy: Directly select the action with the largest value;

[0069] Completely random strategy: Each action has an equal probability of being chosen;

[0070] Greedy strategy: Set a greedy degree ∈ (∈<1). When the generated random number is less than this greedy degree, perform exploration behavior and make a completely random selection; when the generated random number is greater than this greedy degree, perform exploit behavior and choose the action with the largest value.

[0071] Probability distribution strategy: Normalize the results using functions such as Softmax and then sample according to the probability magnitude. For continuous probability distributions, Gaussian distribution is commonly used for probability sampling; this approach can also be considered after discretizing the distribution.

[0072] Furthermore, there are many variations of the algorithm, but they are all basically evolutions based on these basic strategies. They can still be implemented according to the design ideas in this invention and should still fall within the protection scope of this invention.

[0073] Preferably, the reinforcement learning decision-making module includes:

[0074] The input module is used to obtain each action in the reinforcement learning model and its corresponding value, policy type, and accuracy requirements;

[0075] The computation module is used to process the input actions and values ​​according to the strategy type and accuracy requirements to obtain the decision results.

[0076] Preferably, such as Figure 2 As shown, the computational module includes: a data processing module, a random number generation module, a sampling module, a verification module, and an initialization module. The data processing module determines and quantifies the probability values ​​of each action based on the strategy type and precision requirements, and generates an array corresponding to the action probability values. The random number generation module generates random numbers within a preset range. The sampling module samples from the array corresponding to the action probability values ​​based on the random numbers and outputs the action corresponding to the tag of that random number. The verification module verifies the actions output by the sampling module; if the preset conditions are not met, it returns to the random number generation module to regenerate random numbers. The initialization module clears the data processing and verification modules to zero after computation to prevent errors in subsequent calculations.

[0077] Preferably, the data processing module includes a preprocessing module, a quantization module, and an array generation module. The preprocessing module determines the probability value corresponding to each action based on the strategy type; the quantization module quantizes the probability value of each action according to accuracy requirements to obtain a quantization result of a preset width; and the array generation module divides the probability value range into M1 intervals according to accuracy requirements and generates an array corresponding to the action probability value by comparing the interval range of the probability value corresponding to each action.

[0078] Preferably, in the preprocessing module, when determining the probability value corresponding to each action based on the strategy type:

[0079] For deterministic strategies, the probability of the action with the highest value is set to 1, and the probabilities of all other actions are set to 0.

[0080] For a completely random strategy, the probability of each of the N actions is set to 1 / N;

[0081] For the ∈-greedy strategy, the probability of the action with the largest value is set to The probability values ​​of the other (N-1) actions are all set to...

[0082] For strategies that are normalized using the Softmax function and then make decisions based on probability, the probability value corresponding to each action is set to...

[0083] Preferably, in the quantization module, when quantizing the probability values ​​of each action according to the required precision, the minimum precision range that can be represented by the M-bit data width is... The higher the value of M, the higher the precision of the probability it can represent.

[0084] Preferably, the array generation module includes an interval division module, a comparison module, and a sorting and splicing module. The interval division module divides the probability value range into M1 intervals according to the accuracy requirements. The comparison module compares the interval ranges corresponding to the probability values ​​of each action, generating N arrays composed of 0 or 1 signals (each array has a total of M1 signals; the number of 0s and 1s in the array is related to the probability; the higher the probability, the more 1s and fewer 0s in the array). The sorting and splicing module splices the N M1-bit arrays sequentially from low to high to obtain a total array of (N×M1) bits, with corresponding labels from 0 to ((N×M1)-1). For ease of hardware circuit synthesis, the value of (N×M1) is preferably a power of 2.

[0085] Preferably, such as Figure 4 As shown, the random number generation module can generate random numbers in the range of 0 to ((N×M1)-1). (0 to (log2(N×M1)-1)) randomly generated 0s or 1s can represent one random number in the range of 0 to ((N×M1)-1). More preferably, the random number generation module can be connected to any hardware circuit capable of randomly generating 0s and 1s via a port; this embodiment does not impose any restrictions on this.

[0086] Preferably, such as Figure 5As shown, the verification module's main function is to determine whether sampling was successful and to prevent the program from repeatedly failing to sample valid values, leading to lengthy calculations. Specifically, it checks the value at the corresponding position based on the generated random number. If the value is 0, it returns to the random number generation module to regenerate and sample the random number. If the value is 1, it outputs the sample point containing the label corresponding to that random number.

[0087] (0——M1–1) corresponds to sample point A;

[0088] (M1——2M1–1) corresponds to sample point B;

[0089] (2M1——3M1–1) corresponds to sample point C;

[0090] …

[0091] Similarly, although the program is expected to not fail to sample a valid 1 multiple times, some occasional special cases cannot be ruled out. To avoid excessive computation and waste of resources, the verification module also sets up a counter. If no valid value is sampled after Z times, the sample value corresponding to the point sampled in the Zth time is directly output.

[0092] It should be noted that digital circuits are highly programmable and flexible, easy to integrate and automate, and capable of large-scale interconnection. Therefore, this invention uses digital circuits implemented with the hardware description language Verilog. Because digital signals are discrete, there is a loss of precision in the process of expressing input values. At the same time, a balance needs to be struck between precision and computational resources. Therefore, the precision range needs to be clearly defined according to the requirements.

[0093] Example 2

[0094] Based on the reinforcement learning-based data processing system provided in Embodiment 1, this embodiment provides a reinforcement learning-based data processing method, including the following steps:

[0095] 1) Perform reinforcement learning modeling on the data to be processed to determine the action space, its value function, policy type, etc.;

[0096] 2) Obtain each action in the reinforcement learning model and its corresponding value, policy type, and preset accuracy requirements, and use hardware circuits to make rapid decisions to obtain the processing results of the data to be processed.

[0097] Example 3

[0098] This embodiment further illustrates the invention using a simple reinforcement learning scenario with a four-action space, where an agent selects actions using a stochastic strategy.

[0099] like Figure 6 As shown, the agent in the grid can choose from four actions: up, right, down, and left. The agent makes action decisions based on the neural network's calculations and the action selection policy. If it consistently chooses the highest-value action, it may miss some potentially better actions, resulting in low spatial exploration. Sampling actions according to a probability distribution can improve spatial exploration more efficiently during the learning process. Therefore, this invention uses a strategy of decision-making according to a probability distribution as an example. The specific steps for implementing this probability-based sampling circuit include:

[0100] 1) Normalize the sample point values:

[0101] This embodiment uses the Softmax function to convert the values ​​v1, v2, v3, v4 corresponding to each action into probability values ​​P1, P2, P3, P4, that is:

[0102]

[0103]

[0104]

[0105]

[0106] 2) Quantize the probability values: Quantize the probability values ​​according to the required sampling precision.

[0107] In this embodiment, fixed-point decimals are represented using binary, such as... Figure 7 As shown, the probability values ​​in the range of 0-1 are rounded and quantized as follows: 00000,00001,00010,00011,00100,00101,00110,00111,01000,01001,01010,01011,01100,01101,01110,01111,10000,10001,10010,10011,10100,10101,10110,10111,11000,11001,11010,11011,11100,11101,11110,11111.

[0108] 3) Based on the quantized numerical range and the required precision, generate the corresponding array.

[0109] If the probability of a certain action after quantization is between 00000 and 00001, the array generation module generates 00000000000000000;

[0110] If the value is between 00001 and 00011, the array generation module generates 00000000000000001;

[0111] If the value is between 00011 and 00101, the array generation module generates 00000000000000011.

[0112] If the value is between 00101 and 00111, the array generation module generates 00000000000000111;

[0113] If the value is between 00111 and 01001, the array generation module generates 00000000000001111;

[0114] If the value is between 01001 and 01011, the array generation module generates 00000000000011111;

[0115] If the value is between 01011 and 01101, the array generation module generates 00000000000111111;

[0116] If the value is between 01101 and 01111, the array generation module generates 00000000001111111;

[0117] If the value is between 01111 and 10001, the array generation module generates 0000000011111111;

[0118] If the value is between 10001 and 10011, the array generation module generates 0000000111111111;

[0119] If the value is between 10011 and 10101, the array generation module generates 0000001111111111;

[0120] If the value is between 10101 and 10111, the array generation module generates 0000011111111111;

[0121] If the value is between 10111 and 11001, the array generation module generates 0000111111111111;

[0122] If the value is between 11001 and 11011, the array generation module generates 00011111111111111;

[0123] If the value is between 11011 and 11101, the array generation module generates 00111111111111111;

[0124] If the value is between 11011 and 11101, the array generation module generates 01111111111111111;

[0125] If the value is between 11101 and 11111, the array generation module generates 11111111111111111;

[0126] Taking P1 = 0, P2 = 0.25, P3 = 0.25, and P4 = 0.5 as an example, the corresponding generated array is:

[0127] 000000000000000, 000000000001111, 0000000000001111, 00000000011111111

[0128] 4) By splicing the four arrays together, we can obtain the array:

[0129] {000000000000000000000000000001111000000000000111100000000011111111}, where each bit corresponds to a tag from left to right of 0 to (4 × 16 - 1).

[0130] 5) Call the module that generates random numbers (0 or 1) to generate 6-bit random numbers. The range of the random numbers is 0-63(2). 6 -1).

[0131] 6) Call the sampling module, and simultaneously increment COUNT in the counting module. Check if the specific value of the tag position in the array corresponding to the random number is 1. If it is 1, call the output module to output the action of the agent corresponding to the tag. If it is 0, return to the random number generation module and repeat step 5 to generate random numbers. Then sample and count again until the sampling result is 1 or exceeds the maximum value Z specified by the counting module. When the maximum value of the counting module is reached and 1 is still not sampled, the output action is selected as the action corresponding to the Zth sampling.

[0132] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A data processing system based on reinforcement learning, characterized in that... include: The model building module is used to perform reinforcement learning modeling on the data to be processed, and to determine the action space, its value function, and policy type. The reinforcement learning decision module is used to obtain each action in the reinforcement learning model and its corresponding value, policy type and preset accuracy requirements, and to realize fast decision-making using hardware circuits. The output module is used to output the decision results and obtain the processing results of the data to be processed. The reinforcement learning decision module includes: an input module, used to acquire each action in the reinforcement learning model and its corresponding value, policy type, and accuracy requirement; and a calculation module, used to process the input actions and values ​​according to the policy type and accuracy requirement to obtain the decision result. The computational module includes a data processing module, a random number generation module, a sampling module, a verification module, and an initialization module. The data processing module determines and quantifies the probability values ​​of each action based on the strategy type and precision requirements, and generates an array corresponding to the action probability values. The random number generation module generates random numbers within a preset range. The sampling module samples from the array corresponding to the action probability values ​​based on the random numbers and outputs the action corresponding to the tag of the random number. The verification module verifies the actions output by the sampling module; if the preset conditions are not met, it returns to the random number generation module to regenerate random numbers. The initialization module clears the data processing module and the verification module to zero after the calculation is complete. The data processing module includes a preprocessing module, a quantization module, and an array generation module. The preprocessing module is used to determine the probability value corresponding to each action according to the strategy type. The quantization module is used to quantize the probability value of each action according to the accuracy requirements to obtain a quantization result with a preset width. The array generation module is used to divide the probability value range into M1 intervals according to the accuracy requirements, and generate an array corresponding to the action probability value by comparing the interval range of the probability value corresponding to each action. The random number generation module is connected to any hardware circuit capable of randomly generating 0s and 1s via a port, generating... A random number within a given range; The verification module includes a counter module and a judgment module. The counter module is used to count the current number of samples. The judgment module is used to determine whether the sampling was successful and whether the current number of samples has reached a preset threshold. The determination of whether the sampling was successful means: based on the generated random number, find the value at the corresponding position that is 0 or 1. If it is 0, return to the random number generation module, regenerate the random number and sample again. If it is 1, output the action corresponding to the random number. In the preprocessing module, when determining the probability value corresponding to each action based on the strategy type: For deterministic strategies, the probability of the action with the highest value is set to 1, and the probabilities of all other actions are set to 0. For a completely random strategy, the probability of each of the N actions is set to 1 / N; for - Greedy strategy: Set the probability of the action with the highest value to... The probability values ​​for the other (N - 1) actions are all set to... ; For strategies that use the Softmax function for normalization and then make decisions based on probability distributions, the probability value corresponding to each action is set to... ; The array generation module includes an interval division module, a comparison module, and a sorting and splicing module; The interval division module is used to divide the probability value range into intervals according to accuracy requirements. A range; The comparison module is used to compare the range of probability values ​​corresponding to each action and generate N arrays consisting of 0 or 1 signals. The sorting and splicing module is used to sort N groups The array of bits is pieced together sequentially from low to high to obtain... The total array of bits.

2. The data processing system based on reinforcement learning as described in claim 1, characterized in that, The strategy types in the model building module include at least: deterministic strategy, completely random strategy, Greedy strategy or probability distribution strategy.

3. The data processing system based on reinforcement learning as described in claim 1, characterized in that, The random number generation module is connected to any hardware circuit capable of randomly generating 0s and 1s via a port, generating... A random number within a given range.

4. The data processing system based on reinforcement learning as described in claim 1, characterized in that, The verification module includes a counter module and a judgment module. The counter module is used to count the current number of samples. The judgment module is used to determine whether the sampling is successful and whether the current number of samples has reached a preset threshold. The determination of whether the sampling is successful means: based on the generated random number, find the value corresponding to the position as 0 or 1. If it is 0, return to the random number generation module, regenerate the random number and sample. If it is 1, output the action corresponding to the random number.

5. A data processing method based on reinforcement learning, implemented based on the data processing system based on reinforcement learning as described in any one of claims 1 to 4, characterized in that, include: Reinforcement learning modeling is performed on the data to be processed to determine the action space, its value function, and policy type. The algorithm acquires each action in the reinforcement learning model, along with its corresponding value, policy type, and preset accuracy requirements. It then utilizes hardware circuitry to make rapid decisions and obtain the processing results for the data to be processed.

Images