Reinforcement learning based agent dynamic decision network generation method

Abstract

The present application relates to the technical field of artificial intelligence, and more particularly to a method for generating an agent dynamic decision network based on reinforcement learning. The method comprises the following steps: receiving information requirement data input by a user; performing feature analysis on the information requirement data, and extracting business objectives, constraint conditions and key parameters to obtain information requirement analysis results; identifying a task flow corresponding to the information requirement analysis results, and matching an API call chain according to the task flow to obtain a task requirement technical blueprint; generating an agent workflow according to the task requirement technical blueprint and using a preset dynamic workflow engine; performing context context analysis on the language requirement analysis results to obtain context context information; and dividing the agent workflow into super-long thinking chains based on the context context information. In summary, the present application can automatically generate and continuously optimize an agent according to user requirements through reinforcement learning technology, and support efficient decision-making and dynamic adaptation in complex business scenarios.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, and in particular to a reinforcement learning-based agent dynamic decision network generation method. BACKGROUND

[0002] With the continuous innovation of large artificial intelligence model technology, generative artificial intelligence is gradually applied in various fields. However, although the demand for artificial intelligence technology is increasing, there are relatively few agent automation development tools that are easy to use and easy to use for enterprise customers. Most enterprise developers often encounter a series of challenges when trying to integrate artificial intelligence technology into business processes. For example, in the chat machine, intelligent assistant scenario, the model often requires low latency and waiting time; the model is too large to cause memory or memory bottlenecks; it is difficult to optimize the efficiency of large model inference and model accuracy; on the other hand, existing technologies also involve business-level adaptation problems, which are embodied in how to seamlessly integrate intelligent technology with existing business processes, and ensure that the technology application meets the specific needs and compliance requirements of the enterprise. SUMMARY

[0003] Therefore, it is necessary to provide a reinforcement learning-based agent dynamic decision network generation method to solve at least one of the above technical problems.

[0004] To achieve the above purpose, a reinforcement learning-based agent dynamic decision network generation method, the method comprising the following steps:

[0005] Step S1: receiving information demand data input by a user; performing feature analysis on the information demand data, and extracting business targets, constraint conditions and key parameters to obtain information demand analysis results; identifying a task flow corresponding to the information demand analysis results, and matching an API call chain according to the task flow to obtain a task demand technical blueprint;

[0006] Step S2: generating an agent workflow according to the task demand technical blueprint and using a preset dynamic workflow engine; performing context analysis on the language demand analysis results to obtain context information; dividing the agent workflow into super-long thought chains based on the context information; and determining nodes and paths of the agent workflow according to the super-long thought chains;

[0007] Step S3: collecting agent business demand data; dynamically adjusting the nodes and paths in the agent workflow according to the agent business demand data;

[0008] Step S4: Constructing an agent hierarchical reinforcement learning framework, including a strategy layer, a tactic layer, and an execution layer; in the strategy layer, using a reinforcement learning algorithm to predict the long-term business goal achievement path of the agent business demand data and determine the quarterly optimization strategy; in the tactic layer, using a proximal policy optimization algorithm to adjust the agent workflow node configuration and resource allocation scheme in the week dimension; in the execution layer, applying a deep Q network algorithm to optimize the single task execution strategy of the agent workflow in real time;

[0009] Step S5: During the execution of the agent task, constructing and generating an agent dynamic decision network based on the agent hierarchical reinforcement learning framework.

[0010] Preferably, in step S1, the information demand data is analyzed for features, and the business goals, constraints, and key parameters are extracted, including:

[0011] The information demand data is analyzed for features, and the extracted demand features are converted into an information demand vector set;

[0012] The information demand vector set is input into a pre-set intent recognition large model, and an intent recognition matching result is output;

[0013] The intent recognition matching result is divided into business goals, constraints, and key parameters, and its location in the information demand data is recorded to obtain an information demand analysis result.

[0014] Preferably, in step S1, the task flow corresponding to the information demand analysis result is identified, and the API call chain is matched according to the task flow, including:

[0015] The business goals in the information demand analysis result are decomposed into multiple sub-tasks, each of which corresponds to a business operation information;

[0016] Each sub-task is analyzed for features to determine its order and dependency in the task flow;

[0017] According to the order and dependency of the sub-tasks, a task flow graph is constructed, which includes task nodes and task edges, the task nodes representing specific sub-tasks, and the task edges representing the order and dependency between the sub-tasks;

[0018] For each task node in the task flow graph, find the API interface type that matches it;

[0019] According to the task edges in the task flow graph, determine the calling order of the API interfaces to form an API call chain;

[0020] Verify the API call chain to detect whether the input and output of each API interface conform to the task flow order;

[0021] The verified API call chain is bound to the task flowchart to form a complete mapping relationship between the task flowchart and the API call chain, so as to obtain a task demand technology blueprint.

[0022] Preferably, the step S2 of generating the agent workflow according to the task demand technology blueprint and using the preset dynamic workflow engine comprises:

[0023] extracting a starting node and an ending node of the task from the task demand technology blueprint;

[0024] identifying key task nodes in the technology blueprint, each key task node corresponding to a specific business operation decision point, and recording the name and function of each key task node;

[0025] assigning a unique identifier to each key task node and recording its position coordinates in the technology blueprint;

[0026] drawing a preliminary framework diagram of the workflow according to the order and dependency relationship of the key task nodes; wherein the framework diagram contains nodes and connecting lines, the connecting lines representing the execution order between the nodes and using arrows to represent the direction;

[0027] configuring execution parameters for each node in the preliminary framework diagram according to the technology blueprint, the execution parameters including execution time, resource demand and priority;

[0028] instantiating the configured nodes and connecting lines using the preset dynamic workflow engine to generate an executable agent workflow.

[0029] Preferably, the step S2 of dividing the agent workflow into super-long thinking chains based on the context information comprises:

[0030] extracting the context information of each task node from the agent workflow, including the input parameters, output parameters and execution environment of the task node;

[0031] encoding the context information of each task node, and converting the input parameters, output parameters and execution environment into vector form;

[0032] calculating the cosine similarity between the vectors to determine the similarity between adjacent task nodes;

[0033] grouping the task nodes according to the similarity between adjacent task nodes, and grouping the task nodes with a similarity higher than a preset similarity threshold into the same thinking unit;

[0034] for each thinking unit, determining the execution order of the task nodes according to the dependency relationship and priority of the task nodes;

[0035] All thought units are connected in series according to the execution order of the task nodes to form an ultra-long thought chain.

[0036] The ultra-long thought chain is optimized by adjusting the order and path of the task nodes to ensure the coherence and efficiency of the thought chain.

[0037] According to the connection mode of the task nodes and paths in the ultra-long thought chain, the nodes and paths of the intelligent agent workflow are determined.

[0038] Preferably, step S3 includes the following steps:

[0039] Step S31: Collect intelligent agent business demand data from multiple data sources through a pre-set data acquisition interface, including user input, sensor data and historical business records;

[0040] Step S32: Preprocess the intelligent agent business demand data, including data cleaning, format conversion and data labeling, to obtain standard intelligent agent business demand data;

[0041] Step S33: Input the standard intelligent agent business demand data into the dynamic adjustment module to identify the nodes and paths that need to be adjusted in the standard intelligent agent business demand data;

[0042] Step S34: Adjust the task priority, resource demand and execution time of the nodes and paths that need to be adjusted through the dynamic adjustment module, and generate business demand adjustment instructions;

[0043] Step S35: Reconfigure the nodes in the intelligent agent workflow according to the business demand adjustment instructions, including adding, deleting or modifying the attributes of the nodes;

[0044] Step S36: Re-plan the paths in the intelligent agent workflow according to the business demand adjustment instructions, including adjusting the path order and optimizing the path connection mode.

[0045] Preferably, in step S4, at the strategic level, a reinforcement learning algorithm is used to predict the long-term business goal achievement path of the intelligent agent business demand data and determine the quarterly optimization strategy, including the following steps:

[0046] At the strategic level, the task completion rate of the intelligent agent business demand data is determined and recorded as a strategic business indicator;

[0047] Input the strategic business indicator into the reinforcement learning algorithm and initialize the environment state and the initial strategy of the intelligent agent;

[0048] Collect the business execution results of the intelligent agent under different strategies by simulating the interaction between the environment and the intelligent agent, and record the reward value of each strategy;

[0049] According to the business execution result and the reward value of each strategy, the strategy iteration method in the reinforcement learning algorithm is used to update the strategy of the agent;

[0050] Based on the updated strategy, the long-term business target achievement path of the agent business demand data is predicted, and a quarterly optimization strategy is generated.

[0051] Optionally, in step S4, at the tactical level, a proximal policy optimization algorithm is used to adjust the agent workflow node configuration and resource allocation scheme in the week dimension, including:

[0052] At the tactical level, the current node configuration and resource allocation from the agent workflow are extracted; wherein the current node configuration includes the type, number and connection relationship of the node, and the resource allocation includes the type, number and allocation ratio of the resource of each node;

[0053] The current node configuration and resource allocation are input into the proximal policy optimization algorithm, the policy parameters are initialized, and the running environment of the algorithm is set, including the initial state of the simulation environment and the initial strategy of the agent;

[0054] In the week dimension, the resource allocation of each node is evaluated, and the resource utilization efficiency and task completion of each node are recorded to generate resource allocation evaluation results; wherein the resource utilization efficiency includes the type, number and use time of the resource, and the task completion includes the type, number and completion time of the task;

[0055] According to the resource allocation evaluation results, the proximal policy optimization algorithm is used to adjust the resource allocation of each node, optimize the resource utilization efficiency, and update the policy parameters.

[0056] Preferably, in step S4, at the execution level, a deep Q network algorithm is applied to optimize the single task execution strategy of the agent workflow in real time, including:

[0057] At the execution level, the current single task execution strategy from the agent workflow is extracted, including the input parameters, output parameters and execution environment of the task;

[0058] The extracted single task execution strategy is input into the deep Q network algorithm, and the weights and biases of the Q network are initialized;

[0059] At the execution level, the execution process of each task is monitored in real time, and the execution state and result of the task are recorded;

[0060] According to the execution state and result of the task, the deep Q network algorithm is used to calculate the Q value of each task, and the pros and cons of the task execution strategy are evaluated;

[0061] According to the calculated Q value, the execution strategy of the task is adjusted, and the execution efficiency and result of the task are optimized;

[0062] Verify the adjusted execution strategy to ensure it meets business requirements and actual operating conditions;

[0063] Based on the verification results, fine-tune the execution strategy to further optimize the efficiency and results of task execution;

[0064] Apply the fine-tuned execution strategy to the agent workflow to update the single-task execution strategy of the workflow.

[0065] Preferably, step S5 includes the following steps:

[0066] Step S51: Monitor the input and output data of the agent in real time during the execution of the task, and record the intermediate state of the task execution;

[0067] Step S52: Based on the hierarchical reinforcement learning framework, decompose the task execution process into multiple subtasks, each corresponding to a specific decision point;

[0068] Step S53: Assign an independent reinforcement learning module to each subtask, and each module is responsible for optimizing the decision strategy of the subtask;

[0069] Step S54: At each decision point, use the reinforcement learning module to calculate the optimal action under the current state and generate a decision instruction;

[0070] Step S55: Send the generated decision instruction to the agent to guide the next action of the agent;

[0071] Step S56: Update the parameters of the reinforcement learning module according to the execution results of the agent to optimize the decision strategy;

[0072] Step S57: Integrate the decision strategies of all subtasks to build an agent dynamic decision network.

[0073] The unique technical effect of the present application is that by receiving user input information requirement data and performing feature analysis, the business target, constraint condition and key parameter can be accurately extracted, so as to obtain accurate information requirement analysis result. This enables the system to clearly understand the core demand and limitation condition of the user, and provides a solid foundation for subsequent task process identification and API call chain matching. According to the task requirement technical blueprint matched by the task process according to the information requirement analysis result, a clear and accurate technical architecture guidance can be provided for the generation of the agent workflow, ensuring that the generation of the agent workflow is highly consistent with the user demand, and laying a foundation for the efficient operation of the subsequent agent. By using the preset dynamic workflow engine to generate the agent workflow according to the task requirement technical blueprint, a workflow framework suitable for the task requirement can be quickly constructed according to the established technical blueprint. The context context information is obtained by performing context context analysis on the language requirement analysis result, and then the agent workflow is divided into super long thinking chain based on this, and the node and path of the agent workflow are determined. This process can fully consider the context correlation of the language requirement, so that the structure of the agent workflow is more reasonable and the logic is more coherent, effectively improving the work efficiency and accuracy of the agent when processing complex tasks. The node and path in the agent workflow are dynamically adjusted according to the collected agent business requirement data, realizing the dynamic adaptability of the agent workflow. This process can optimize and adjust the workflow in a timely manner according to the changes of the actual business requirement, ensuring that the agent workflow is always consistent with the business requirement, thereby improving the flexibility and adaptability of the agent when facing different business scenarios, and enhancing the business processing capability of the agent. The constructed agent hierarchical reinforcement learning framework includes a strategy layer, a tactical layer and an execution layer, which can optimize the workflow of the agent from different dimensions. At the strategy layer, a reinforcement learning algorithm is used to predict the long-term business target achievement path of the agent business requirement data and determine the quarterly optimization strategy, so that the agent can grasp the long-term direction of business development from a macro perspective, plan the optimal path in advance, and ensure that the agent maintains efficient and stable development in the long-term business development process. At the tactical layer, a near-end strategy optimization algorithm is used to adjust the node configuration and resource allocation scheme of the agent workflow in a week dimension, which can adjust the node configuration and resource allocation of the workflow in a timely manner according to the changes of the short-term business requirement, improve the resource utilization efficiency and task execution effect of the agent in the short-term business execution process. At the execution layer, a deep Q network algorithm is applied to real-time optimize the single task execution strategy of the agent workflow, so that the agent can quickly and accurately make optimal decisions when executing specific tasks, improving the execution quality and efficiency of single tasks. During the execution of the agent task, the agent dynamic decision network is constructed and generated based on the agent hierarchical reinforcement learning framework, which enables the agent to make decisions in real time during the task execution process based on the hierarchical reinforcement learning framework.The construction of this dynamic decision-making network enables intelligent agents to make quick and accurate decisions that adapt to the current task requirements based on real-time business demand data and workflow status, further improving the decision-making ability and task execution efficiency of intelligent agents in complex business environments, and ensuring that intelligent agents can complete various tasks efficiently and stably. Attached Figure Description

[0074] Figure 1 This is a flowchart illustrating the steps of a method for generating dynamic decision networks for intelligent agents based on reinforcement learning.

[0075] Figure 2 for Figure 1 A detailed flowchart illustrating the implementation steps of step S3.

[0076] Figure 3 for Figure 1 A detailed flowchart illustrating the implementation steps of step S5.

[0077] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0078] The technical method of the present invention will now be clearly and completely described 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 embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0079] Furthermore, the accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor methods and / or microcontroller methods.

[0080] It should be understood that although the terms "first," "second," etc., may be used herein to describe various units, these units should not be limited by these terms. These terms are used merely to distinguish one unit from another. For example, without departing from the scope of the exemplary embodiments, a first unit may be referred to as a second unit, and similarly, a second unit may be referred to as a first unit. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0081] To achieve the above objectives, please refer toFigures 1 to 3 A reinforcement learning-based agent dynamic decision network generation method, the method comprising the following steps:

[0082] Step S1: receiving user input information requirement data; performing feature analysis on the information requirement data, and extracting business targets, constraint conditions and key parameters to obtain information requirement analysis results; identifying the task flow corresponding to the information requirement analysis results, and matching API call chains according to the task flow to obtain a task demand technical blueprint;

[0083] Step S2: generating an agent workflow according to the task demand technical blueprint and using a preset dynamic workflow engine; performing context context analysis on the language requirement analysis results to obtain context context information; dividing the agent workflow into super-long thinking chains based on the context context information, and determining the nodes and paths of the agent workflow according to the super-long thinking chains;

[0084] Step S3: collecting agent business requirement data; dynamically adjusting the nodes and paths in the agent workflow according to the agent business requirement data;

[0085] Step S4: constructing an agent hierarchical reinforcement learning framework, including a strategic layer, a tactical layer and an execution layer; in the strategic layer, using a reinforcement learning algorithm to predict the long-term business target achievement path of the agent business requirement data and determine a quarterly optimization strategy; in the tactical layer, using a proximal policy optimization algorithm to adjust the agent workflow node configuration and resource allocation scheme in a week dimension; in the execution layer, applying a deep Q network algorithm to optimize the single task execution strategy of the agent workflow in real time;

[0086] Step S5: during the execution of the agent, constructing and generating an agent dynamic decision network based on the agent hierarchical reinforcement learning framework.

[0087] In the embodiment of the application, reference Figure 1 is shown, which is a step flowchart of a reinforcement learning-based agent dynamic decision network generation method of the application. In this example, the reinforcement learning-based agent dynamic decision network generation method comprises the following steps:

[0088] Step S1: receiving user input information requirement data; performing feature analysis on the information requirement data, and extracting business targets, constraint conditions and key parameters to obtain information requirement analysis results; identifying the task flow corresponding to the information requirement analysis results, and matching API call chains according to the task flow to obtain a task demand technical blueprint;

[0089] In the embodiment of the present application, the system receives the information requirement data input by the user through the natural language interface and records it completely in the system log. The BERT model is used for intent recognition and entity extraction of the natural language text input by the user. The BERT model encodes the input text by using its pre-trained weight parameters to extract the word embedding vector in the text. The system sets a threshold parameter, and when the similarity of the word embedding vector and the vector of the predefined business target, constraint condition and key parameter exceeds the threshold (for example, 0.8), it is considered that the corresponding feature is successfully extracted. Based on the generation ability of GPT, the business target, constraint condition and key parameter obtained by the semantic analysis layer are converted into a technical blueprint containing task flow, API call chain and evaluation index. The GPT model generates detailed technical specification text according to the input structured information, combined with its pre-trained parameters and generation logic. The system sets a keyword matching threshold, for example, 80% of the keyword matching degree. When the keyword in the technical specification matches the keyword in the task flow template with a degree exceeding the threshold, it is considered that the matching is successful. The system further analyzes the logical relationship of the keywords that match successfully by using a logical reasoning algorithm to ensure that the generated task flow conforms to the business logic. According to each link in the task flow, the pre-defined API interface library is called to generate a specific API call chain; according to the key parameters in the technical specification, each API interface in the API call chain is configured with parameters. The system stores the task flow and API call chain in a structured manner in the database to form a task demand technical blueprint. The format of the technical blueprint is JSON format, which contains each step of the task flow, detailed information of the API call chain, and evaluation indexes. The system stores the generated technical blueprint in the task blueprint library for subsequent use by the agent construction and optimization module. The storage path of the technical blueprint is “ / task_blueprint / {task_id}.json”, wherein {task_id} is the unique identifier of the task.

[0090] Step S2: generating an agent workflow according to the task demand technical blueprint and using a preset dynamic workflow engine; performing context context analysis on the language demand analysis result to obtain context context information; dividing the agent workflow into super-long thinking chains based on the context context information, and determining the nodes and paths of the agent workflow according to the super-long thinking chains;

[0091] In the embodiment of the application, when the dynamic workflow engine is initialized, the task flow and API call chain information in the task requirement technical blueprint are loaded. The engine sets the maximum number of nodes of the workflow to 128 to support the generation of super-long thinking chains. The engine converts each step and its corresponding API call chain into a workflow node according to the step order in the task flow. Each node contains a node ID, a node type (such as API call, decision point, result output, etc.), and input and output parameters of the node. The context analysis module extracts semantic information from the natural language text input by the user, and generates context information by combining the business objectives, constraints and key parameters in the task requirement technical blueprint. For example, if the information requirement input by the user is "construct a supply chain risk early warning system", the context information includes keywords such as "supply chain management", "risk assessment" and "real-time monitoring". The module analyzes the extracted context information to determine the specific position and role of each keyword in the task flow. For example, the "supply chain management" keyword is related to the demand prediction and supplier negotiation steps, and the "risk assessment" keyword is related to the demand prediction and logistics scheduling steps. The system divides the agent workflow into multiple sub-chains according to the keywords and logical relationships in the context information. Each sub-chain contains a series of related nodes, forming a logically coherent thinking chain. For example, the super-long thinking chain of the supply chain risk early warning system is divided into a "demand prediction sub-chain" (node 1-node 2) and a "logistics scheduling sub-chain" (node 2-node 3). The system determines the nodes and paths in each sub-chain according to the logical relationships in the context information. For example, in the "demand prediction sub-chain", the output of node 1 (demand prediction) is used as the input of node 2 (supplier negotiation), forming a path from node 1 to node 2. In the "logistics scheduling sub-chain", the output of node 2 (supplier negotiation) is used as the input of node 3 (logistics scheduling), forming a path from node 2 to node 3.

[0092] Step S3: Collecting agent business requirement data; dynamically adjusting nodes and paths in the agent workflow according to the agent business requirement data;

[0093] In the embodiments of the present application, the data collection module configures data source access parameters, for example, for business system logs, set the log file path as " / var / log / business_system.log", for user interaction records, set the database connection parameters as "host:port:database:user:password". According to business requirements, set the data collection frequency. For example, for business data with high real-time requirements, set the collection frequency to once every 5 minutes; for log data, set the collection frequency to once every 30 minutes. The collected data is preprocessed by the preprocessing module to clean and format the data. The preprocessing module removes noise and redundant information in the data to ensure the accuracy and consistency of the data. For example, for user interaction records, the preprocessing module will remove duplicate records and convert the data to a unified JSON format. The agent workflow adjustment module parses the collected business requirement data and extracts key information such as changes in business objectives, new constraint conditions, etc. For example, if the business requirement data contains the information "add new supplier evaluation index", the change requirement is extracted. According to the parsed business requirement data, the agent workflow adjustment module dynamically adjusts the nodes in the workflow. For example, if a new supplier evaluation index needs to be added, a new evaluation index sub-node is added in the supplier negotiation node (node 2). The adjustment module updates the input and output parameters of the node to ensure the compatibility of the new node with the existing node. The agent workflow adjustment module dynamically adjusts the path in the workflow according to the business requirement data. For example, if the new business requirement requires adding a supplier credit evaluation step after supplier negotiation, a new supplier credit evaluation node (node 2.1) is inserted between the supplier negotiation node (node 2) and the logistics scheduling node (node 3), and the path is updated to node 2→node 2.1→node 3. When adjusting the nodes and paths, the agent workflow adjustment module will dynamically configure the parameters of the related nodes according to the business requirement data. For example, for the newly added supplier credit evaluation node (node 2.1), configure the evaluation index parameters such as credit score threshold, evaluation period, etc.

[0094] Step S4: Construct an agent hierarchical reinforcement learning framework, including a strategic layer, a tactical layer and an execution layer; in the strategic layer, use a reinforcement learning algorithm to predict the long-term business goal achievement path of the agent business requirement data and determine the quarterly optimization strategy; in the tactical layer, use a proximal policy optimization algorithm to adjust the agent workflow node configuration and resource allocation scheme in the week dimension; in the execution layer, apply a deep Q network algorithm to real-time optimize the single task execution strategy of the agent workflow;

[0095] In the embodiment of the present application, the system calls a model-based reinforcement learning (MBRL) algorithm module at the strategic layer. The module loads historical business data and execution results of the agent workflow to build a world model. The world model describes the transition probability of the agent between different states and the reward value in different states through a state transition probability matrix and a reward function. The system sets the dimension of the state transition probability matrix to n x n, where n is the number of states, and the value range of the reward function is . Using the MBRL algorithm combined with Monte Carlo tree search (MCTS) technology, a large number of paths are simulated, and the optimal path is selected as the prediction result of the long-term business goal achievement path. According to the prediction result, the system generates a quarterly optimization strategy, including task allocation, resource allocation, and workflow optimization within the quarter. The system calls a proximal policy optimization (PPO) algorithm module at the tactical layer. The module initializes the policy network with a hidden layer size of 256 neurons, a learning rate of 0.001, and a discount factor γ of 0.99. Based on the execution results of the agent workflow and business demand data, the PPO algorithm calculates the policy gradient, updates the policy network parameters, and dynamically adjusts the workflow node configuration and resource allocation scheme. For example, according to the weekly business data, the resource allocation proportion of each node in the workflow is adjusted to optimize task execution efficiency. According to the optimization results of the PPO algorithm, the system adjusts the resource allocation scheme, such as dynamically adjusting the allocation proportion of CPU, memory, and other resources according to the priority and resource demand of the task to ensure efficient use of resources. The system calls a deep Q-network (DQN) algorithm module at the execution layer. The module initializes the Q-network with a hidden layer size of 128 neurons, a learning rate of 0.0005, and a discount factor γ of 0.95. Through the DQN algorithm, the value function of the state-action pair is learned. The DQN algorithm uses the experience replay (Experience Replay) technology to store and replay the interaction experience of the agent, optimizing the Q-network parameters. For example, according to the execution results of the agent in different states, the weights of the Q-network are updated to optimize the single-task execution strategy. According to the optimization results of the DQN algorithm, the system adjusts the single-task execution strategy of the agent workflow in real time. For example, according to the current state of the task, the optimal action is selected to optimize the execution efficiency and response time of the task.

[0096] Step S5: During the execution of the agent task, an agent dynamic decision network is built and generated based on the agent hierarchical reinforcement learning framework.

[0097] In the embodiments of the present application, first, real-time data generated by the agent during task execution is obtained from the task execution monitoring module. These data include key information such as task progress, resource consumption, and environmental feedback. These data are recorded and stored in the task execution log database. The table structure of the database is pre-designed and includes fields such as task ID, timestamp, progress percentage, resource usage, and environmental feedback type, ensuring data integrity and traceability. Subsequently, the system calls the decision network construction module, which constructs the agent's dynamic decision network based on the three levels of the hierarchical reinforcement learning framework: the strategic level, the tactical level, and the execution level. At the strategic level, the system uses the previously predicted long-term business goal achievement path and quarterly optimization strategy, combined with real-time data from the current task execution, to update the agent's long-term strategy through the policy iteration technique in the reinforcement learning algorithm. Specifically, the system sets the convergence threshold of policy iteration to 0.01. When the policy change between two consecutive iterations is less than the threshold, the policy is considered to have converged, resulting in a stable long-term strategy. This strategy will serve as the top-level guidance for the decision network, determining the agent's behavior direction at the macro level. At the tactical level, the system adjusts the agent's short-term strategy using the proximal policy optimization algorithm (PPO) based on the weekly dimension of the agent's workflow node configuration and resource allocation scheme, as well as real-time task execution data. The system sets the parameters of the PPO algorithm, including a learning rate of 0.001, a discount factor γ of 0.99, and a batch size of 64. Through these parameters, the PPO algorithm can dynamically adjust the transition strategy between nodes and the resource allocation strategy based on the agent's current workflow node state and resource usage, to improve the agent's task execution efficiency and resource utilization efficiency in the short term. For example, if the task execution progress of a node lags behind, the PPO algorithm will adjust resource allocation, increase the computing resources of that node, and optimize the transition strategy between nodes to reduce unnecessary waiting time. At the execution level, the system optimizes the agent's immediate behavior based on the deep Q-network (DQN) algorithm. The DQN algorithm module receives real-time data from the task execution monitoring module, including the current task state, available actions, and immediate rewards. The system sets the parameters of the DQN algorithm, such as a learning rate of 0.0005, a discount factor γ of 0.95, and an experience replay pool size of 10,000. The DQN algorithm continuously updates Q values by learning the value function of state-action pairs, combined with the experience replay technique, to select the optimal action for the agent at each time step. For example, in a specific task execution scenario, the agent faces multiple selectable actions. The DQN algorithm will select the action with the highest expected reward based on the current state and estimated Q values to achieve efficient task completion. Finally, the system integrates the strategies of the strategic level, the tactical level, and the execution level to form a complete agent dynamic decision network.The decision net is stored in the form of a graph in a decision net database, each node representing a decision point, an edge representing a transition from one decision point to another, and the weight of the edge representing the probability or priority of the transition. The table structure of the decision net database is reasonably designed and can store information such as the topological structure of the decision net, node attributes, and edge attributes. In the process of the agent executing a task, the system queries the decision net database in real time, dynamically adjusts the behavior of the agent according to the current state and the guidance of the decision net, ensures that the agent can efficiently and flexibly cope with various complex situations, and achieves the completion of the task goal.

[0098] Preferably, the feature analysis of the information requirement data in step S1 includes extracting the business target, constraint condition and key parameter, and the feature analysis of the information requirement data in step S1 includes extracting the business target, constraint condition and key parameter.

[0099] The extracted requirement features are converted into an information requirement vector set;

[0100] The information requirement vector set is input into a preset intent recognition large model, and an intent recognition matching result is output;

[0101] The intent recognition matching result is divided into a business target, a constraint condition and a key parameter, and the positions thereof in the information requirement data are recorded to obtain an information requirement analysis result.

[0102] In the embodiment of the present application, after the system receives the information requirement data in the form of natural language input by the user, a feature extraction module is called. This module uses the word embedding method in natural language processing technology, such as Word2Vec or BERT, to extract features from the information requirement data. Taking BERT as an example, the system takes the information requirement data as input, and through the encoder part of the BERT model, each word or phrase is converted into a fixed-dimensional vector. In specific operation, the system sets the parameters of the BERT model, such as the hidden layer dimension of 768 and the maximum sequence length of 512. In this way, each word or phrase in the information requirement data is converted into a 768-dimensional vector to form a set of information requirement vectors. At the same time, the system records the position information of each vector in the original information requirement data for subsequent processing. The system inputs the generated set of information requirement vectors into a pre-set large intent recognition model. This large model is based on the Transformer architecture and processes the input vector set through multiple layers of self-attention mechanism and feedforward neural network. In the processing process, the system sets the parameters of the large model, such as the number of layers is 12, the number of attention heads is 12, and the hidden layer dimension is 768. The large model performs deep learning and pattern recognition on the set of information requirement vectors according to these parameters, and outputs the intent recognition matching result. The output result is in the form of probability distribution, indicating the matching degree of the information requirement data with each type of pre-defined intent. For example, if the information requirement data is "construct a supply chain risk early warning system", the large model will output a probability of 0.9 for matching with the "system construction" intent and a probability of 0.7 for matching with the "risk assessment" intent. After receiving the intent recognition matching result output by the large model, the system calls an intent analysis module. This module divides the intent recognition matching result into business objectives, constraint conditions and key parameters according to pre-defined intent classification rules. In specific operation, the system sets an intent classification threshold, for example 0.5, and divides the intent recognition results with a matching probability higher than the threshold into the corresponding categories. For business objectives, the system extracts results matching with intents such as "system construction" and "function implementation"; for constraint conditions, it extracts results matching with intents such as "time limit" and "cost control"; for key parameters, it extracts results matching with intents such as "technical requirements" and "performance indicators". At the same time, the system determines the specific positions of these categories in the original information requirement data according to the previously recorded position information. Finally, the system integrates the divided business objectives, constraint conditions and key parameters and their position information into the information requirement analysis result, and stores it in the database in a structured form, providing a basis for subsequent agent construction and optimization.

[0103] Preferably, the step S1 of identifying the information requirement analysis result corresponding to the task flow comprises:

[0104] The business target in the information requirement analysis result is decomposed into a plurality of subtasks, each of which corresponds to a business operation information;

[0105] Each subtask is subjected to feature recognition to determine its sequence and dependency in the task flow;

[0106] According to the sequence and dependency of the subtasks, a task flow chart is constructed, which includes task nodes and task edges, the task nodes representing specific subtasks and the task edges representing the sequence and dependency between the subtasks;

[0107] For each task node in the task flow chart, an API interface type matching the task node is found;

[0108] According to the task edges in the task flow chart, the calling sequence between the API interfaces is determined to form an API calling chain;

[0109] The API calling chain is verified to detect whether the input and output of each API interface conform to the task flow sequence;

[0110] The verified API calling chain is bound to the task flow chart to form a complete mapping relationship between the task flow and the API calling chain, thereby obtaining a task requirement technical blueprint.

[0111] In this embodiment of the invention, business objective fields, such as "building a supply chain risk early warning system," are extracted from the information requirements parsing results. Based on the business process specifications in the domain knowledge base, the business objective is decomposed into multiple sub-tasks. For example, "building a supply chain risk early warning system" is decomposed into sub-tasks such as "demand forecasting," "supplier evaluation," "risk analysis," and "early warning generation." The domain knowledge base predefines the mapping relationship between business objectives and sub-tasks, and the system decomposes them according to this mapping relationship. The decomposed sub-tasks are stored in a task decomposition table, whose structure includes fields such as sub-task ID, sub-task name, and sub-task description. For example, the sub-task with sub-task ID 1 is named "demand forecasting" and described as "forecasting supply chain demand." Keywords are extracted from the description of each sub-task using the TF-IDF algorithm. For example, the keywords "demand" and "forecasting" are extracted for the "demand forecasting" sub-task. Using RAG retrieval technology, semantic analysis is performed on the sub-task descriptions to calculate the semantic similarity between sub-tasks. A semantic similarity threshold of 0.7 is set; when the semantic similarity between two sub-tasks exceeds this threshold, they are considered to have a dependency relationship. Based on keyword matching and RAG search results, the dependencies between subtasks are determined. For example, "demand forecasting" depends on the completion of "data collection," and "supplier evaluation" depends on the results of "demand forecasting." These dependencies are stored in a dependency table, with fields including subtask ID, dependent subtask ID, and dependency type. A task node is created for each subtask, assigned a unique node ID. For example, the node ID for subtask "data collection" is 1, and the node ID for "demand forecasting" is 2. Based on the dependency table, task edges are created to represent the order and dependencies between subtasks. For example, creating task edge 1→2 indicates that "demand forecasting" will be executed after "data collection" is completed. The task flowchart is stored graphically in a task flow database, with tables including fields such as node ID, node name, edge ID, edge start point, and edge end point. For example, the node table records node ID and node name, and the edge table records edge ID, edge start point, and edge end point. Based on the subtask name and description, a matching API interface type is searched in a predefined API interface library. The API interface library contains information such as the API interface name, function description, input parameters, and output parameters. For example, the API interface matched for the "Data Collection" subtask is "DataCollector". The matching results are stored in an API matching table, whose structure includes fields such as subtask ID, API interface name, and API interface description. For example, the API interface name for subtask ID 1 is "DataCollector", and its description is "Collect supply chain related data". The order in which the API interfaces are called is determined based on the task edges in the task flow diagram. For example, based on the task flow... Figure 1→ 2 → 3, and the API call chain is "DataCollector → DemandPredictor → SupplierEvaluator". The API call chain is stored in the API call chain table, and the table structure includes task ID, API interface name, call order, etc. For example, the API interface name of task ID 1 is "DataCollector", and the call order is 1. According to the input and output parameter definition in the API interface library, it is verified whether the input and output of each API interface meet the task flow order. For example, it is verified whether the output of "DataCollector" is the input of "DemandPredictor", and whether the output of "DemandPredictor" is the input of "SupplierEvaluator". The verification rules are set, such as input and output parameter type matching, data format consistency, etc. The verification result is output in the form of a Boolean value. If the verification is passed, the subsequent operation is continued; if the verification fails, an error message is returned and the API call chain is prompted to be adjusted again. Each task node is associated with the corresponding API interface to form a complete mapping relationship between the task flow and the API call chain. For example, task node 1 is bound to API interface "DataCollector", and task node 2 is bound to API interface "DemandPredictor". The binding result is stored in the task demand technology blueprint table, and the table structure includes task node ID, API interface name, input and output parameters, etc. For example, the API interface name of task node ID 1 is "DataCollector", the input parameter is "supply chain data", and the output parameter is "demand prediction data". The finally generated task demand technology blueprint is stored in JSON format, which contains detailed information of the task flow chart and API call chain, and provides a basis for subsequent agent construction and optimization.

[0112] Preferably, step S2 of generating an agent workflow according to the task demand technology blueprint and using a preset dynamic workflow engine includes:

[0113] extracting the starting node and the ending node of the task from the task demand technology blueprint;

[0114] identifying key task nodes in the technology blueprint, each key task node corresponding to a specific business operation decision point, and recording the name and function of each key task node;

[0115] assigning a unique identifier to each key task node and recording its position coordinates in the technology blueprint;

[0116] According to the order and dependency relationship of the key task nodes, a preliminary framework diagram of the workflow is drawn; wherein the framework diagram contains nodes and connecting lines, the connecting lines represent the execution order between the nodes, and the direction is indicated by arrows;

[0117] According to the technical blueprint, an execution parameter of each node in the preliminary framework diagram is configured, the execution parameter including execution time, resource requirement and priority;

[0118] The configured nodes and connecting lines are instantiated in a preset dynamic workflow engine to generate an executable agent workflow.

[0119] In the embodiment of the present application, the system call technology blueprint analysis module reads the data stored in the task demand technology blueprint table. The technology blueprint table includes task node ID, node name, node type, input and output parameters, etc. The system identifies the starting node through the node type field. The starting node is usually marked as "START" type. For example, the node ID is 1, the node name is "data collection", and the node type is "START". The system identifies the end node through the node type field. The end node is usually marked as "END" type. For example, the node ID is 10, the node name is "result output", and the node type is "END". The information of the starting node and the end node is stored in the workflow configuration table, and the table structure includes node ID, node name, node type, position coordinates, etc. The key node identification module is called to identify the key task nodes through the node function description field. The key nodes usually include decision logic or important business operations. For example, the node ID is 3, the node name is "demand prediction", and the function description is "predict future demand according to historical data". The name and function of the key task nodes are recorded in the workflow configuration table. For example, the node ID is 3, the node name is "demand prediction", and the function description is "predict future demand according to historical data". The unique identifier generation module is called to generate a unique identifier for each key task node. For example, the node ID is 3, and the unique identifier "UUID-001" is generated. According to the layout information in the technology blueprint, the position coordinates of each key task node are recorded. For example, the node ID is 3, and the position coordinates are (100, 200). The unique identifier and the position coordinates are updated to the workflow configuration table. For example, the node ID is 3, the unique identifier is "UUID-001", and the position coordinates are (100, 200). The workflow framework drawing module is called to initialize an empty framework. The framework is stored in a graphical form, including nodes and connection lines. According to the information in the workflow configuration table, each key task node is added to the framework. The node is represented in a graphical form, including node name, unique identifier and position coordinates. According to the dependency relationship in the technology blueprint, the connection line is drawn to represent the execution order between nodes. The connection line uses an arrow to represent the direction. For example, the node ID is 1, the node "data collection" is connected to the node ID 2, the node "demand prediction", and a arrow from (100, 100) to (200, 200) is drawn. The preliminary framework is stored in the workflow framework database, and the database table structure includes node ID, node name, unique identifier, position coordinates, connection line start point, connection line end point, etc. The execution parameters of each node are extracted from the technology blueprint table. For example, the node ID is 3, the node "demand prediction" has an execution time of 10 minutes, a resource requirement of CPU 2 cores and memory 4GB, and a high priority. The parameter configuration module is called to configure the extracted execution parameters to the workflow configuration table.For example, the node with node ID 3, "demand prediction", has an execution time of 10 minutes, resource requirements of CPU 2 cores and memory 4 GB, and a high priority. Update the configured execution parameters to the workflow configuration table. For example, the node with node ID 3, "demand prediction", has an execution time of 10 minutes, resource requirements of CPU 2 cores and memory 4 GB, and a high priority. Call the dynamic workflow engine to initialize the workflow instantiation module. The workflow engine supports multiple task scheduling algorithms, such as priority scheduling, resource scheduling, etc. According to the information in the workflow configuration table, each node is instantiated as a task object. The task object contains node name, unique identifier, execution parameters, etc. For example, the node with node ID 3, "demand prediction", is instantiated as a task object, containing a unique identifier "UUID-001", an execution time of 10 minutes, resource requirements of CPU 2 cores and memory 4 GB, and a high priority. According to the connection line information in the workflow configuration table, the connection line is instantiated as the dependency relationship between tasks. The dependency relationship is stored in the form of reference to the task object. For example, task object "data collection" depends on task object "demand prediction". Assemble the instantiated nodes and connection lines into a complete agent workflow. The workflow is stored in a graphical form and supports real-time monitoring and dynamic adjustment. Store the generated agent workflow in the workflow instance database, and the database table structure includes task object ID, task name, unique identifier, execution parameters, dependency relationship, etc. For example, the task object ID is 3, the task name is "demand prediction", the unique identifier is "UUID-001", the execution time is 10 minutes, the resource requirements are CPU 2 cores and memory 4 GB, the priority is high, and the dependency relationship is task object ID 1 "data collection".

[0120] Preferably, the agent workflow is divided into super-long thinking chains based on the context information in step S2, and the nodes and paths of the agent workflow are determined according to the super-long thinking chains, including:

[0121] Extract the context information of each task node from the agent workflow, including the input parameters, output parameters and execution environment of the task node;

[0122] Encode the context information of each task node, and convert the input parameters, output parameters and execution environment into vector form;

[0123] Calculate the cosine similarity between vectors to determine the similarity between adjacent task nodes;

[0124] Group the task nodes according to the similarity between adjacent task nodes, and group the task nodes with similarity higher than the preset similarity threshold into the same thinking unit;

[0125] For each thinking unit, determine the execution order of the task nodes according to the dependency and priority of the task nodes;

[0126] Connect all thinking units in the order of task node execution to form a super-long thinking chain;

[0127] Optimize the super-long thinking chain by adjusting the order and path of the task nodes to ensure the coherence and efficiency of the thinking chain;

[0128] Determine the nodes and paths of the agent workflow according to the connection mode of the task nodes and paths in the super-long thinking chain.

[0129] In the embodiment of the present application, the system call context context extraction module reads the task node information stored in the workflow configuration table. The workflow configuration table includes task node ID, node name, input parameter, output parameter, execution environment and other fields. For example, the node name of task node ID 3 is "demand prediction", the input parameter is "historical sales data", the output parameter is "future demand prediction", and the execution environment is "CPU 2 core, memory 4GB". The extracted context context information is stored in the context context information table, and the table structure includes task node ID, input parameter, output parameter, execution environment and other fields. For example, the input parameter of task node ID 3 is "historical sales data", the output parameter is "future demand prediction", and the execution environment is "CPU 2 core, memory 4GB". The system call vector encoding module converts the context context information into a vector form using a predefined encoding rule. For example, the input parameter "historical sales data" is converted into the vector 1,0,0, the output parameter "future demand prediction" is converted into the vector 0,1,0, and the execution environment "CPU 2 core, memory 4GB" is converted into the vector 0,0,1. The encoded vector is stored in the context vector table, and the table structure includes task node ID, input parameter vector, output parameter vector, execution environment vector and other fields. For example, the input parameter vector of task node ID 3 is 1,0,0, the output parameter vector is 0,1,0, and the execution environment vector is 0,0,1. The system call cosine similarity calculation module calculates the cosine similarity between the input parameter vector, the output parameter vector and the execution environment vector of each task node. The cosine similarity is determined by calculating the ratio of the dot product of two vectors to the product of their modules. For example, the cosine similarity of the input parameter vector 1,0,0 of task node ID 3 and the input parameter vector 1,0,0 of task node ID 4 is 1. The calculated cosine similarity is stored in the similarity table, and the table structure includes task node ID1, task node ID2, similarity and other fields. For example, the similarity of task node ID 3 and task node ID 4 is 1. The system sets the similarity threshold to 0.8. When the cosine similarity between two task nodes is higher than the threshold, they are classified into the same thinking unit. The system call grouping module groups the task nodes according to the data in the similarity table. For example, the cosine similarity of task node ID 3 and task node ID 4 is 1, which is higher than the threshold 0.8, so they are classified into the same thinking unit. The grouping result is stored in the grouping information table, and the table structure includes thinking unit ID, task node ID and other fields. For example, thinking unit ID 1 includes task node ID 3 and task node ID 4. The system extracts the dependency relationship and priority of each task node from the workflow configuration table. For example, the priority of task node ID 3 is high, and it depends on task node ID 2. The system call execution order determination module determines the execution order of the task nodes according to the dependency relationship and priority.For example, the priority of the task node with ID 2 is medium, the priority of the task node with ID 3 is high, and the task node with ID 3 depends on the task node with ID 2, so the execution order is 2→3. The determined execution order is stored in an execution order table, and the table structure includes fields such as thought unit ID, task node ID, and execution order. For example, the task node execution order of the thought unit with ID 1 is 2→3. The system calls the concatenation module, and concatenates all thought units according to the data in the execution order table. For example, the task node execution order of the thought unit with ID 1 is 2→3, the task node execution order of the thought unit with ID 2 is 4→5, and the concatenated super-long thought chain is 2→3→4→5. The super-long thought chain is stored in a super-long thought chain table, and the table structure includes fields such as task node ID and execution order. For example, the super-long thought chain table records that the execution order of the task node with ID 2 is 1, the execution order of the task node with ID 3 is 2, the execution order of the task node with ID 4 is 3, and the execution order of the task node with ID 5 is 4. The system sets optimization rules, such as reducing the waiting time between task nodes and optimizing resource utilization. The system calls the optimization module, and adjusts the order and path of the task nodes according to the optimization rules. For example, by adjusting the order of the task nodes, the waiting time between the task nodes is reduced, and the resource utilization efficiency is improved. The optimized super-long thought chain is stored in an optimized super-long thought chain table, and the table structure includes fields such as task node ID and optimized execution order. For example, the optimized super-long thought chain table records that the optimized execution order of the task node with ID 2 is 1, the optimized execution order of the task node with ID 3 is 2, the optimized execution order of the task node with ID 4 is 3, and the optimized execution order of the task node with ID 5 is 4. The system calls the node and path determination module, and determines the nodes and paths of the agent workflow according to the data in the optimized super-long thought chain table. For example, the node name of the task node with ID 2 is “data collection”, the node name of the task node with ID 3 is “demand prediction”, and the path is 2→3. The determined nodes and paths are updated to the agent workflow to form the final agent workflow. The final agent workflow is stored in a graphical form, and supports real-time monitoring and dynamic adjustment. The final agent workflow is stored in a workflow instance database, and the database table structure includes fields such as task node ID, task name, execution order, and dependency relationship. For example, the task name of the task node with ID 2 is “data collection”, the execution order is 1, and the dependency relationship is none; the task name of the task node with ID 3 is “demand prediction”, the execution order is 2, and the dependency relationship is the task node with ID 2.

[0130] As an example of the present application, reference is made to Fig. 1, which shows a flowchart of the method according to the present application. In this example, the step S3 comprises: Figure 2

[0131] ​Step S31: Collecting agent business requirement data from multiple data sources through a pre-set data collection interface, including user input, sensor data, and historical business records;

[0132] Step S32: Preprocessing the agent business requirement data, including data cleaning, format conversion, and data labeling, to obtain standard agent business requirement data;

[0133] Step S33: Inputting the standard agent business requirement data into the dynamic adjustment module to identify the nodes and paths that need to be adjusted in the standard agent business requirement data;

[0134] Step S34: Dynamically adjusting the task priority, resource demand, and execution time of the nodes and paths that need to be adjusted through the dynamic adjustment module, and generating business requirement adjustment instructions;

[0135] Step S35: Reconfiguring the nodes in the agent workflow according to the business requirement adjustment instructions, including adding, deleting, or modifying the attributes of the nodes;

[0136] Step S36: Replanning the paths in the agent workflow according to the business requirement adjustment instructions, including adjusting the path order and optimizing the path connection method.

[0137] In the embodiments of the present application, the system is configured with a data collection interface that supports the access of multiple data sources. For example, user input is collected through a Web interface, sensor data is collected through an Internet of Things (IoT) interface, and historical business records are collected through a database interface. The system receives user input data in real time through the Web interface, such as user input of business requirements in a Web form. The system receives sensor data through the IoT interface, such as real-time data from temperature sensors, pressure sensors, and the like. The system reads historical business records from an enterprise database through the database interface, such as past sales data, inventory records, and the like. The collected data is stored in a pre-set data warehouse that supports large-scale data storage and fast querying. The system calls a data cleaning module to remove noise and redundant information from the data. For example, duplicate user input records are removed, and outliers in sensor data are filtered out. The system calls a format conversion module to convert data in different formats to a uniform format. For example, JSON format data input by the user is converted to XML format, and binary format sensor data is converted to text format. The system calls a data labeling module to label the data for subsequent processing. For example, business requirement data input by the user is labeled as "high priority" or "low priority", and sensor data is labeled as "normal" or "abnormal". The pre-processed data is stored in a pre-processing data table, which includes fields such as data ID, data content, data format, and data label. The pre-processed data is input into a dynamic adjustment module, which reads and analyzes the data. The dynamic adjustment module identifies nodes and paths that need to be adjusted in the data through a pre-set rule engine. For example, if a business requirement labeled as "high priority" in the data does not match a certain node in the existing workflow, the node is identified as a node that needs to be adjusted. The identification result is stored in an adjustment requirement table, which includes fields such as data ID, node ID that needs to be adjusted, and path ID that needs to be adjusted. The dynamic adjustment module adjusts the task priority of the relevant node according to the priority label in the business requirement data. For example, the priority of a node labeled as "high priority" is adjusted from "medium" to "high". The dynamic adjustment module adjusts the resource requirement of the relevant node according to the resource requirement label in the business requirement data. For example, the CPU resource requirement of a certain node is increased from 2 cores to 4 cores. The dynamic adjustment module adjusts the execution time of the relevant node according to the execution time label in the business requirement data. For example, the execution time of a certain node is adjusted from 10 minutes to 15 minutes. The dynamic adjustment module generates business requirement adjustment instructions, which include information such as node ID that needs to be adjusted, adjusted task priority, resource requirement, and execution time. The generated adjustment instructions are stored in an adjustment instruction table, which includes fields such as instruction ID, node ID, adjusted task priority, resource requirement, and execution time. The system calls a workflow configuration module to reconfigure the nodes in the workflow according to the adjustment instructions.For example, add a new node according to the adjustment instruction, delete a node that is no longer needed, or modify the task priority, resource requirement, and execution time of the existing node. Store the reconfigured node information in the workflow configuration table, and the table structure includes node ID, node name, task priority, resource requirement, execution time, etc. Call the workflow path planning module, and re-plan the path in the workflow according to the adjustment instruction. For example, adjust the path order to adapt to the new task priority, and optimize the path connection mode to reduce the waiting time between nodes. Store the re-planned path information in the workflow path table, and the table structure includes path ID, start node ID, end node ID, path order, etc.

[0138] Preferably, in step S4, at the strategic layer, the long-term business goal achievement path of the agent business demand data is predicted and the quarterly optimization strategy is determined by using a reinforcement learning algorithm, including the following steps:

[0139] At the strategic layer, the task completion rate of the agent business demand data is measured and recorded as a strategic business indicator;

[0140] The strategic business indicator is input into the reinforcement learning algorithm, and the initial strategy of the agent and the environment state are initialized;

[0141] By simulating the interaction between the environment and the agent, the business execution results of the agent under different strategies are collected, and the reward value of each strategy is recorded;

[0142] According to the business execution results and the reward value of each strategy, the strategy iteration method in the reinforcement learning algorithm is used to update the strategy of the agent;

[0143] Based on the updated strategy, the long-term business goal achievement path of the agent business demand data is predicted, and a quarterly optimization strategy is generated.

[0144] In the embodiment of the present application, the system call task completion rate measurement module calculates the task completion rate according to the task execution record in the intelligent agent business demand data. The task completion rate is determined by the ratio of the number of completed tasks to the total number of tasks. For example, if the total number of tasks is 100 and the number of completed tasks is 80, the task completion rate is 80%. The task completion rate is used as a strategic business indicator and stored in the strategic business indicator table. The strategic business indicator table includes fields such as indicator ID, indicator name, and indicator value. For example, the indicator name of indicator ID 1 is "task completion rate" and the indicator value is 80%. The system selects a model-based reinforcement learning (MBRL) algorithm as the optimization algorithm of the strategic layer. The system initializes the environment state, including the current task completion rate, resource usage, time progress, and other information. For example, the current task completion rate is 80%, the resource usage rate is 60%, and the time progress is the second quarter. The system initializes the initial policy of the agent, which is generated based on historical data and preset rules. For example, the initial policy stipulates that resource input should be increased when the task completion rate is less than 70% and resource input should be reduced when the task completion rate is higher than 90%. The system sets up a simulation environment, which includes a task generator, a resource allocator, and a time controller. The task generator generates tasks according to business demand, the resource allocator allocates resources according to the agent's policy, and the time controller controls the simulation time progress. The system runs the simulation environment, and the agent interacts with the environment according to the initial policy. For example, in the first interaction, the agent selects to increase resource input according to the initial policy, and the task completion rate increases to 85% and the resource usage rate increases to 70%. The system records the reward value of each policy, which is calculated based on the task completion rate, resource usage rate, and time progress. For example, the reward value calculation formula is: reward value = task completion rate × 0.5 + resource usage rate × 0.3 + time progress × 0.2. In the first interaction, the reward value is 0.85 × 0.5 + 0.70 × 0.3 + 0.5 × 0.2 = 0.735. The system sets the parameters of policy iteration, including learning rate, discount factor, and iteration times. For example, the learning rate is 0.01, the discount factor is 0.99, and the iteration times are 1000. The system runs the policy iteration algorithm and updates the agent's policy according to the reward value of each interaction. For example, in the first iteration, the agent adjusts the policy according to the reward value 0.735, increasing the resource input proportion when the task completion rate is less than 75%. The system judges whether the policy converges or not, and the convergence condition is that the policy changes less than 0.001 in 10 consecutive iterations. If the policy converges, stop iteration; otherwise, continue the next iteration. The system uses the updated policy combined with Monte Carlo tree search (MCTS) technology to predict the long-term business goal achievement path of the intelligent agent business demand data. For example, it is predicted that through policy adjustment in each quarter in the next four quarters, the task completion rate will gradually increase to 95%.The system generates a quarterly optimization strategy according to the predicted long-term business target achievement path. The quarterly optimization strategy includes task allocation, resource allocation, and strategy adjustment suggestions for each quarter. For example, the first quarter suggests increasing resource input to improve task completion rate, and the second quarter suggests optimizing task allocation to improve resource utilization efficiency. The generated quarterly optimization strategy is stored in the optimization strategy table, and the table structure includes fields such as strategy ID, quarter, task allocation, resource allocation, and strategy adjustment suggestions. For example, the strategy ID is 1, the quarter is the first quarter, the task allocation is "increase resource input for task X", the resource allocation is "increase CPU resources by 2 cores", and the strategy adjustment suggestion is "increase resource input when task completion rate is less than 75%".

[0145] Preferably, in step S4, at the tactical level, a proximal policy optimization algorithm is used to adjust the agent workflow node configuration and resource allocation scheme in the weekly dimension, including:

[0146] At the tactical level, the current node configuration and resource allocation from the agent workflow are extracted; the current node configuration includes the type, number, and connection relationship of the nodes, and the resource allocation includes the type, number, and allocation proportion of each node's resources;

[0147] The current node configuration and resource allocation are input into the proximal policy optimization algorithm, the policy parameters are initialized, and the running environment of the algorithm is set, including the initial state of the simulation environment and the initial strategy of the agent;

[0148] In the weekly dimension, the resource allocation of each node is evaluated, and the resource usage efficiency and task completion of each node are recorded to generate resource allocation evaluation results; the resource usage efficiency includes the type, number, and usage time of the resources, and the task completion includes the type, number, and completion time of the tasks;

[0149] According to the resource allocation evaluation results, the proximal policy optimization algorithm is used to adjust the resource allocation of each node, optimize the resource usage efficiency, and update the policy parameters.

[0150] In the embodiment of the present application, the system calls the node configuration extraction module to read the node configuration information in the agent workflow. The node configuration information includes the type, number and connection relationship of the nodes. For example, the workflow contains three types of nodes: data processing nodes, analysis nodes and output nodes, with quantities of 5, 3 and 2 respectively, and the connection relationship is that the data processing nodes are connected to the analysis nodes, and the analysis nodes are connected to the output nodes. The system calls the resource allocation extraction module to read the resource allocation of each node. The resource allocation includes the resource type, quantity and allocation ratio. For example, the data processing node is allocated CPU resources of 2 cores and memory of 4 GB, with an allocation ratio of 50%; the analysis node is allocated CPU resources of 4 cores and memory of 8 GB, with an allocation ratio of 30%; and the output node is allocated CPU resources of 2 cores and memory of 4 GB, with an allocation ratio of 20%. The extracted node configuration and resource allocation are stored in the node configuration table and the resource allocation table. The node configuration table contains node ID, node type, quantity, connection relationship and other fields; the resource allocation table contains node ID, resource type, quantity, allocation ratio and other fields. The system selects the proximal policy optimization (PPO) algorithm as the optimization algorithm of the tactical layer. The system initializes the strategy parameters of the PPO algorithm, including the learning rate, discount factor, batch size and the like. For example, the learning rate is 0.001, the discount factor is 0.99, and the batch size is 64. The system sets the running environment of the algorithm, including the initial state of the simulation environment and the initial strategy of the agent. The initial state of the simulation environment includes the current node configuration and resource allocation, and the initial strategy of the agent is generated based on historical data and preset rules. For example, the initial strategy stipulates that resource allocation is increased when resource utilization efficiency is less than 60%, and resource allocation is reduced when task completion rate is higher than 80%. The system sets the evaluation period to be once a week. The system calls the resource utilization efficiency evaluation module to calculate the resource utilization efficiency of each node. The resource utilization efficiency includes the type, quantity and use time of the resource. For example, the CPU resource utilization efficiency of the data processing node is 70%, the memory resource utilization efficiency is 65%, and the use time is 40 hours. The system calls the task completion condition evaluation module to record the task completion condition of each node. The task completion condition includes the type, quantity and completion time of the task. For example, the data processing node has completed 10 data cleaning tasks, with a completion time of 30 hours; the analysis node has completed 5 data analysis tasks, with a completion time of 20 hours. The resource utilization efficiency and task completion condition are stored in the resource allocation evaluation table, which includes node ID, resource type, resource utilization efficiency, task type, task quantity, completion time and other fields. The system sets the adjustment strategy of the PPO algorithm, including the resource adjustment threshold and the task completion rate threshold. For example, the resource adjustment threshold is 60%, and the task completion rate threshold is 80%. The system runs the PPO algorithm to adjust the resource allocation of each node according to the resource allocation evaluation result.For example, the resource utilization efficiency of the data processing node is 70%, which is higher than the resource adjustment threshold of 60%, so its resource allocation ratio is increased from 50% to 60%; the task completion rate of the analysis node is 75%, which is lower than the task completion rate threshold of 80%, so its resource allocation ratio is reduced from 30% to 25%. The system updates the policy parameters of the PPO algorithm according to the adjusted resource allocation. For example, the updated learning rate is 0.0008, the discount factor is 0.98, and the batch size is 128. The adjusted resource allocation and the updated policy parameters are stored in the resource allocation adjustment table and the policy parameter table. The resource allocation adjustment table contains fields such as node ID, resource type, adjusted quantity, and adjusted allocation ratio; the policy parameter table contains fields such as parameter name and parameter value.

[0151] Preferably, in step S4, the deep Q network algorithm is applied in the execution layer to optimize the single-task execution strategy of the agent workflow in real time, which includes:

[0152] In the execution layer, the current single-task execution strategy is extracted from the agent workflow, including the input parameters, output parameters, and execution environment of the task;

[0153] The extracted single-task execution strategy is input into the deep Q network algorithm, and the weights and biases of the Q network are initialized;

[0154] In the execution layer, the execution process of each task is monitored in real time, and the execution status and results of the task are recorded;

[0155] According to the execution status and results of the task, the Q value of each task is calculated using the deep Q network algorithm to evaluate the pros and cons of the task execution strategy;

[0156] According to the calculated Q value, the execution strategy of the task is adjusted to optimize the execution efficiency and results of the task;

[0157] The adjusted execution strategy is verified to ensure that it meets the business requirements and actual operating conditions;

[0158] According to the verification result, the execution strategy is fine-tuned to further optimize the execution efficiency and results of the task;

[0159] The fine-tuned execution strategy is applied to the agent workflow to update the single-task execution strategy of the workflow.

[0160] In the embodiments of the present application, the system call strategy adjustment module adjusts the execution strategy of the task according to the Q value. For example, the Q value of the task "data cleaning" with task ID 1 is 0.885, which is lower than the preset Q value threshold 0.9, so the execution strategy needs to be adjusted. The adjustment strategy includes increasing resource allocation, optimizing task execution process, etc. For example, the CPU resource allocation of the task "data cleaning" is increased from 2 cores to 4 cores, and the task execution process is optimized to reduce the execution time. The system updates the execution strategy of the task according to the adjusted strategy. For example, the execution strategy of the task "data cleaning" with task ID 1 is updated as: the input parameter is "original data file", the output parameter is "cleaned data file", and the execution environment is "CPU 4 cores, memory 4 GB". The adjusted execution strategy is stored in the execution strategy table, and the table structure includes task ID, task name, input parameter, output parameter, execution environment, etc. For example, the task name of the task with task ID 1 is "data cleaning", the input parameter is "original data file", the output parameter is "cleaned data file", and the execution environment is "CPU 4 cores, memory 4 GB". The system configuration verification module verifies whether the adjusted execution strategy meets the business requirements and actual operation conditions. The verification module simulates the execution of the task to check whether the execution result of the task meets the expectation. The system runs the verification module to simulate the execution of the task "data cleaning" with task ID 1 and checks the execution result of the task. For example, the execution result of the task "data cleaning" is successful, the output parameter is "cleaned data file", the task completion time is 2024-10-01 09:00, and the task execution result meets the expectation. The verification result is stored in the verification result table, and the table structure includes task ID and verification result. For example, the verification result of the task with task ID 1 is successful. The system configuration fine-tuning module fine-tunes the execution strategy according to the verification result. The fine-tuning module further optimizes the execution efficiency and result of the task by adjusting the resource allocation and execution process of the task. The system runs the fine-tuning module to fine-tune the task "data cleaning" with task ID 1. For example, the fine-tuning module adjusts the CPU resource allocation of the task "data cleaning" from 4 cores to 3 cores, and optimizes the task execution process to reduce resource waste. The fine-tuned execution strategy is stored in the execution strategy table, and the table structure includes task ID, task name, input parameter, output parameter, execution environment, etc. For example, the task name of the task with task ID 1 is "data cleaning", the input parameter is "original data file", the output parameter is "cleaned data file", and the execution environment is "CPU 3 cores, memory 4 GB". The system calls the workflow update module to apply the fine-tuned execution strategy to the agent workflow. For example, the execution strategy of the task "data cleaning" with task ID 1 is updated as: the input parameter is "original data file", the output parameter is "cleaned data file", and the execution environment is "CPU 3 cores, memory 4 GB".The updated execution strategy is stored in a workflow configuration table, and the table structure includes task ID, task name, input parameter, output parameter, execution environment and the like. For example, the task name of the task with task ID 1 is "data cleaning", the input parameter is "original data file", the output parameter is "cleaned data file", and the execution environment is "CPU 3 core, memory 4 GB".

[0161] As an example of the present application, reference is made to Fig. 1, which shows a schematic diagram of a system for training an agent according to the present application. Figure 3 In this example, the step S5 includes:

[0162] Step S51: Real-time monitoring of the input and output data of the agent during task execution, recording the intermediate state of task execution;

[0163] Step S52: Based on the hierarchical reinforcement learning framework, the task execution process is divided into multiple sub-tasks, each sub-task corresponding to a specific decision point;

[0164] Step S53: Assign an independent reinforcement learning module to each sub-task, and each module is responsible for optimizing the decision strategy of the sub-task;

[0165] Step S54: At each decision point, use the reinforcement learning module to calculate the optimal action under the current state and generate a decision instruction;

[0166] Step S55: Send the generated decision instruction to the agent to guide the next action of the agent;

[0167] Step S56: Update the parameters of the reinforcement learning module according to the execution result of the agent, and optimize the decision strategy;

[0168] Step S57: Integrate the decision strategies of all sub-tasks to build an agent dynamic decision network.

[0169] In the embodiments of the present application, the system is configured with a real-time monitoring module to monitor the input and output data of the intelligent agent. The monitoring module records the execution status of the task through the task execution log, including the start time, end time, execution progress, resource usage, etc. The system records the input and output data of the intelligent agent in real time. For example, the input data of the task "data cleaning" with task ID 1 is "original data file", and the output data is "cleaned data file". The system records the intermediate state of task execution. For example, during the execution of the task "data cleaning" with task ID 1, the intermediate state includes data reading completion, data preprocessing completion, data cleaning completion, etc. The input and output data and intermediate state of the task are stored in the task execution log table, and the table structure includes task ID, task name, input data, output data, intermediate state, etc. For example, the task name of the task with task ID 1 is "data cleaning", the input data is "original data file", the output data is "cleaned data file", and the intermediate state includes data reading completion, data preprocessing completion, and data cleaning completion. The system is configured with a task decomposition module, which decomposes the task execution process into multiple sub-tasks according to predefined business rules and domain knowledge base. For example, the task "data cleaning" is decomposed into three sub-tasks: "data reading", "data preprocessing", and "data cleaning". Each sub-task corresponds to a specific decision point. For example, the decision point of the "data reading" sub-task is to select which data reading method, the decision point of the "data preprocessing" sub-task is to select which preprocessing algorithm, and the decision point of the "data cleaning" sub-task is to select which cleaning strategy. The decomposed sub-tasks are stored in the sub-task table, and the table structure includes sub-task ID, sub-task name, decision point, etc. For example, the sub-task name of the sub-task with sub-task ID 1 is "data reading", and the decision point is "select data reading method". The system configures an independent reinforcement learning module for each sub-task. For example, a reinforcement learning module is configured for the "data reading" sub-task, and another reinforcement learning module is configured for the "data preprocessing" sub-task. The system initializes the parameters of each reinforcement learning module, including learning rate, discount factor, experience replay pool size, etc. For example, the learning rate is 0.001, the discount factor is 0.99, and the experience replay pool size is 10000. The configuration information of each reinforcement learning module is stored in the reinforcement learning module table, and the table structure includes module ID, sub-task ID, learning rate, discount factor, experience replay pool size, etc. For example, the sub-task ID of the module with module ID 1 is 1, the learning rate is 0.001, the discount factor is 0.99, and the experience replay pool size is 10000. The system perceives the current state of task execution. For example, at the decision point of the "data reading" sub-task, the current state includes the format of the input data, the size of the data, etc. The system calls the reinforcement learning module to calculate the optimal action according to the current state. For example, at the decision point of the "data reading" sub-task, according to the format and size of the input data, the most suitable data reading method is selected.The system generates a decision instruction to guide the next action of the agent. For example, the generated decision instruction is "select CSV format data reading method". The generated decision instruction is stored in the decision instruction table, and the table structure includes fields such as decision point ID and decision instruction. For example, the decision instruction of the decision point ID 1 is "select CSV format data reading method". The system sends the decision instruction to the agent through the preset communication interface. The communication interface supports real-time data transmission to ensure that the instruction can be delivered in time. The agent receives the decision instruction and executes the next action according to the instruction. For example, the agent receives the instruction "select CSV format data reading method", and starts to execute the data reading operation. The system records the result of the agent executing the decision instruction. For example, the agent successfully executes the data reading operation, and the execution result is recorded as success. The system evaluates the result of the agent executing the decision instruction. For example, the execution efficiency and resource usage of the data reading operation are evaluated. The system updates the parameters of the reinforcement learning module according to the execution result using the parameter update method in the reinforcement learning algorithm. For example, the learning rate, discount factor and other parameters are adjusted according to the execution result to optimize the decision strategy. The updated parameters are stored in the reinforcement learning module table, and the table structure includes fields such as module ID, subtask ID, learning rate, discount factor and experience replay pool size. For example, the subtask ID of the module ID 1 is 1, the updated learning rate is 0.0008, the discount factor is 0.98, and the experience replay pool size is 12000. The system integrates the decision strategy of each subtask into a complete decision network. For example, the decision strategies of the three subtasks "data reading", "data preprocessing" and "data cleaning" are integrated into a dynamic decision network. The system constructs the dynamic decision network of the agent, and the decision network is stored in a graphical form, including nodes and edges. The node represents the subtask, and the edge represents the execution order and dependency relationship between the subtasks. The constructed dynamic decision network of the agent is stored in the decision network database, and the database table structure includes fields such as node ID, node name, edge start point and edge end point. For example, the node ID 1 is "data reading", the edge start point is 1, and the edge end point is 2, indicating that the "data preprocessing" subtask is executed after the "data reading" subtask is completed.

[0170] Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting, the scope of the application being defined by the appended claims and not by the above description, and all variations falling within the meaning and the scope of the equivalent elements of the application file are intended to be included in the present application.

[0171] The foregoing is considered as illustrative only of the principles of the application. Numerous modifications and changes will readily occur to those skilled in the art, and it is intended to embrace all such modifications and changes that fall within the scope of the application. Accordingly, the application is not to be restricted in scope to the specific embodiments disclosed herein but is to be accorded the full scope that the principles and novel features request appropriately granted.

Claims

1. A method for generating a reinforcement learning-based intelligent agent dynamic decision network, characterized in that, The method comprises the following steps: Step S1: receiving user input information requirement data; performing feature analysis on the information requirement data, and extracting business objectives, constraint conditions and key parameters to obtain information requirement analysis results; Identify the task flow corresponding to the information requirement analysis results, and match the API call chain according to the task flow to obtain a task requirement technology blueprint; Step S2: generating an agent workflow according to the task requirement technology blueprint and using a preset dynamic workflow engine; performing context context analysis on the language requirement analysis results to obtain context context information; Divide the agent workflow into super-long thought chains based on the context context information, and determine the nodes and paths of the agent workflow according to the super-long thought chains; The method of dividing the agent workflow into super-long thought chains based on the context context information and determining the nodes and paths of the agent workflow according to the super-long thought chains comprises: Extract the context context information of each task node from the agent workflow, including the input parameters, output parameters and execution environment of the task node; Encode the context context information of each task node, and convert the input parameters, output parameters and execution environment into vector form; Calculate the cosine similarity between vectors to determine the similarity between adjacent task nodes; Group the task nodes according to the similarity between adjacent task nodes, and group the task nodes with a similarity higher than a preset similarity threshold into the same thought unit; For each thought unit, determine the execution order of the task nodes according to the dependency relationship and priority of the task nodes; Serially connect all thought units according to the execution order of the task nodes to form a super-long thought chain; Optimize the super-long thought chain by adjusting the order and path of the task nodes to ensure the coherence and efficiency of the thought chain; Determine the nodes and paths of the agent workflow according to the connection mode of the task nodes and paths in the super-long thought chain; Step S3: collecting agent business requirement data; dynamically adjusting the nodes and paths in the agent workflow according to the agent business requirement data; Step S4: building an agent hierarchical reinforcement learning framework, including a strategy layer, a tactic layer and an execution layer; in the strategy layer, using a reinforcement learning algorithm to predict the long-term business goal achievement path of the agent business requirement data and determine the quarterly optimization strategy; in the tactic layer, using a proximal policy optimization algorithm to adjust the agent workflow node configuration and resource allocation scheme in the week dimension; in the execution layer, applying a deep Q network algorithm to optimize the single task execution strategy of the agent workflow in real time; The method of using a reinforcement learning algorithm in the strategy layer to predict the long-term business goal achievement path of the agent business requirement data and determine the quarterly optimization strategy comprises: In the strategy layer, measure the task completion rate of the agent business requirement data and record it as a strategic business indicator; Input the strategic business indicator into the reinforcement learning algorithm, and initialize the environment state and the initial strategy of the agent; Collect the business execution results of the agent under different strategies by simulating the interaction between the environment and the agent, and record the reward value of each strategy; According to the business execution results and the reward value of each strategy, update the strategy of the agent using the policy iteration method in the reinforcement learning algorithm; Based on the updated strategy, the long-term business target achievement path of the agent business demand data is predicted, and a quarterly level optimization strategy is generated; At the tactical level, the proximal policy optimization algorithm is used to adjust the agent workflow node configuration and resource allocation scheme in the week dimension, which includes: At the tactical level, the current node configuration and resource allocation in the agent workflow are extracted, wherein the current node configuration includes the type, number and connection relationship of the node, and the resource allocation includes the type, number and allocation ratio of the resources of each node; The current node configuration and resource allocation are input into the proximal policy optimization algorithm, the policy parameters are initialized, and the running environment of the algorithm is set, including the initial state of the simulation environment and the initial strategy of the agent; In the week dimension, the resource allocation of each node is evaluated, and the resource usage efficiency and task completion of each node are recorded to generate resource allocation evaluation results; wherein the resource usage efficiency includes the type, number and usage time of the resources, and the task completion includes the type, number and completion time of the tasks; According to the resource allocation evaluation results, the proximal policy optimization algorithm is used to adjust the resource allocation of each node, optimize the resource usage efficiency, and update the policy parameters; Step S5: During the execution of the agent task, an agent dynamic decision network is constructed and generated based on the agent hierarchical reinforcement learning framework.

2. The method of claim 1, wherein, In step S1, the information demand data is analyzed and the business target, constraint condition and key parameter are extracted, which includes: The information demand data is feature extracted, and the extracted demand features are converted into an information demand vector set; The information demand vector set is input into the preset intent recognition large model, and the intent recognition matching result is output; The intent recognition matching result is divided into business target, constraint condition and key parameter, and its position in the information demand data is recorded to obtain the information demand analysis result.

3. The method of claim 2, wherein, In step S1, the task flow corresponding to the information demand analysis result is identified, and the API calling chain is matched according to the task flow, which includes: The business target in the information demand analysis result is decomposed into multiple subtasks, wherein each subtask corresponds to a business operation information; The characteristics of each subtask are identified to determine its order and dependency in the task flow; According to the order and dependency of the subtasks, a task flow diagram is constructed, which includes task nodes and task edges, the task nodes represent specific subtasks, and the task edges represent the order and dependency between the subtasks; For each task node in the task flow diagram, find the API interface type matched therewith; According to the task edge in the task flow diagram, the calling order of the API interfaces is determined to form an API calling chain; The API calling chain is verified to detect whether the input and output of each API interface meet the task flow order; The verified API calling chain is bound with the task flow diagram to form a complete mapping relationship between the task flow and the API calling chain, thereby obtaining a task demand technology blueprint.

4. The method of claim 1, wherein, In step S2, the agent workflow is generated based on the task demand technology blueprint and using the preset dynamic workflow engine, which includes: The starting node and the ending node of the task are extracted from the task demand technology blueprint; Identify key task nodes in the technology blueprint, each corresponding to a specific business operation decision point, and record the name and function of each key task node; Assign a unique identifier to each key task node and record its location coordinates in the technology blueprint; According to the order and dependency of the key task nodes, draw a preliminary framework diagram of the workflow; The framework diagram contains nodes and connecting lines, and the connecting lines represent the execution order between nodes, and the direction is indicated by arrows; Configure execution parameters for each node in the preliminary framework diagram according to the technology blueprint, including execution time, resource requirements and priority; Use the preset dynamic workflow engine to instantiate the configured nodes and connecting lines to generate an executable agent workflow.

5. The method of claim 1, wherein, Step S3 includes the following steps: Step S31: Collect agent business demand data from multiple data sources through a preset data collection interface, including user input, sensor data and historical business records; Step S32: Preprocess the agent business demand data, including data cleaning, format conversion and data labeling, to obtain standard agent business demand data; Step S33: Input the standard agent business demand data into the dynamic adjustment module, and use the dynamic adjustment module to identify the nodes and paths that need to be adjusted in the standard agent business demand data; Step S34: Adjust the task priority, resource demand and execution time of the nodes and paths that need to be adjusted through the dynamic adjustment module, and generate business demand adjustment instructions; Step S35: Reconfigure the nodes in the agent workflow according to the business demand adjustment instructions, including adding, deleting or modifying the attributes of the nodes; Step S36: Replan the paths in the agent workflow according to the business demand adjustment instructions, including adjusting the path order and optimizing the path connection method.

6. The method of claim 1, wherein, In step S4, in the execution layer, the deep Q network algorithm is applied to optimize the single task execution strategy of the agent workflow, including: In the execution layer, extract the current single task execution strategy from the agent workflow, including the input parameters, output parameters and execution environment of the task; Input the extracted single task execution strategy into the deep Q network algorithm, initialize the weights and biases of the Q network; In the execution layer, real-time monitor the execution process of each task, record the execution state and result of the task; According to the execution state and result of the task, calculate the Q value of each task using the deep Q network algorithm to evaluate the pros and cons of the task execution strategy; According to the calculated Q value, adjust the execution strategy of the task to optimize the execution efficiency and result of the task; Verify the adjusted execution strategy to ensure that it meets the business demand and actual operation conditions; According to the verification result, fine-tune the execution strategy to further optimize the execution efficiency and result of the task; Apply the fine-tuned execution strategy to the agent workflow to update the single task execution strategy of the workflow.

7. The method of claim 1, wherein, Step S5 includes the following steps: Step S51: In the process of executing the task of the agent, real-time monitor the input and output data of the agent, and record the intermediate state of the task execution; Step S52: Based on the hierarchical reinforcement learning framework, the task execution process is decomposed into multiple sub-tasks, each corresponding to a specific decision point; Step S53: Assign an independent reinforcement learning module to each sub-task, each module responsible for optimizing the decision strategy of the sub-task; Step S54: At each decision point, use the reinforcement learning module to calculate the optimal action under the current state and generate a decision instruction; Step S55: Send the generated decision instruction to the agent to guide the next action of the agent; Step S56: According to the execution result of the agent, update the parameters of the reinforcement learning module and optimize the decision strategy; Step S57: Integrate the decision strategies of all sub-tasks to build an agent dynamic decision network.

Images