Method and system for autonomous deduction and decision model construction of agent based on decision tree

Abstract

This invention discloses a method and system for constructing an autonomous inference and decision-making model for intelligent agents based on decision trees, belonging to the field of intelligent decision-making and autonomous planning technology. The method constructs a state vector based on environmental topology, observation data, and task constraints; combines rule-feasible candidate actions and learned supplementary candidate actions to construct a decision tree; utilizes cost distribution to predict the mean cost and uncertainty of the output candidate actions from the network; and constructs a joint score by combining event intensity, action feasibility, and risk coefficient to achieve risk constraint tree search; in the replanning mode, only branches affected by events are locally reconstructed according to a smoothed amplitude-limited frequency, and the prediction network, risk coefficient, and pruning margin are updated online based on the execution residuals. This invention can improve the efficiency, stability, and robustness of autonomous decision-making.

Description

Technical Field

[0001] This invention relates to the field of intelligent decision-making and autonomous planning technology, and in particular to a method and system for constructing an intelligent agent's autonomous inference and decision-making model for dynamic environments and multi-constraint task scenarios. This model is based on state modeling, dual-source generation of candidate actions, cost distribution prediction, risk-constrained decision tree search, event-induced replanning, local reconstruction of affected branches, and residual closed-loop correction. Background Technology

[0002] In applications such as robot navigation, autonomous action planning for unmanned systems, task scheduling in complex scenarios, emergency evacuation simulations, and multi-agent collaborative decision-making, agents typically need to combine environmental topology, real-time observation information, and task constraints to continuously output the optimal action or action sequence for the current moment. Existing technologies often employ graph search, decision tree search, rule-based planning, sampling planning, and cost evaluation methods combined with machine learning models to achieve autonomous simulation and online decision-making.

[0003] As application scenarios evolve from static, low-constraint environments to dynamic, highly perturbed, and multi-constraint environments, existing autonomous decision-making methods have gradually revealed several shortcomings in engineering applications. Firstly, in terms of candidate action generation, existing technologies typically favor a single candidate source: one approach relies primarily on rule-based filtering to generate candidate actions. While this can ensure actions meet constraints to some extent, overly stringent filtering can lead to insufficient candidate action coverage, resulting in the omission of potentially optimal actions. Another approach relies mainly on learning models to directly provide action recommendations. Although this approach has strong exploratory capabilities, it easily introduces actions that do not meet constraints such as safe distance, time window, capacity, or resource availability, leading to an increase in invalid branches, higher filtering costs, and decreased decision reliability in subsequent search phases.

[0004] Secondly, regarding cost assessment and tree search coupling, existing technologies mostly use single-point cost estimation as the basis for decision tree expansion and pruning, lacking characterization of prediction uncertainty and a unified scoring mechanism for action feasibility. When environmental obstacles, capacity, tasks, or sample distribution shift, relying solely on single-point cost for branch selection can easily lead to high-risk branches being mistakenly retained and potentially preferred branches being prematurely pruned, thus affecting the robustness of tree search and the stability of decision results.

[0005] Secondly, regarding environmental event-triggered replanning, existing technologies often employ single-threshold triggering or fixed update frequencies to control the replanning process. When the event intensity fluctuates around the threshold, the system is prone to frequently entering and exiting the replanning state, causing jitter in action output, fluctuations in computational load, and repeated switching of the execution layer. Furthermore, some existing solutions directly reconstruct the entire decision tree after triggering replanning, making it difficult to distinguish between truly affected local branches and unaffected stable branches, leading to unnecessary increases in computational overhead and impacting system real-time performance.

[0006] Furthermore, in terms of online adaptive capabilities, existing technologies typically lack a closed-loop correction mechanism based on the difference between actual and predicted costs after action execution. They fail to utilize execution feedback to continuously correct cost prediction models, risk coefficients, and pruning parameters, causing prediction errors and search biases to gradually accumulate during continuous decision-making, thereby affecting the long-term stability and adaptability of the system under dynamic environments, task changes, and distribution drift conditions.

[0007] Therefore, there is an urgent need for a method and system for constructing autonomous inference and decision-making models for intelligent agents based on decision trees, so as to take into account the feasibility and coverage of candidate actions, search efficiency and decision robustness, event triggering stability and real-time replanning under dynamic environments and multiple constraints, and have online closed-loop self-correction capabilities based on execution feedback. Summary of the Invention

[0008] The purpose of this invention is to provide a method and system for constructing an intelligent agent's autonomous inference and decision model based on decision trees, in order to solve the technical problems in the prior art, such as the single method of candidate action generation, the difficulty in balancing action feasibility and coverage, the weak coupling between cost evaluation and tree search, the easy jitter of event-triggered replanning, the high computational load of full reconstruction, and the lack of online closed-loop correction driven by execution feedback.

[0009] To achieve the above objectives, this invention constructs a state vector based on environmental topology, observation data, and task constraints. Based on this, it combines a set of rule-feasible candidate actions with a set of learned supplementary candidate actions to construct a candidate action set and a decision tree. The cost distribution prediction network outputs the ensemble cost mean and prediction uncertainty corresponding to each candidate action, and combines event intensity, action feasibility, and risk coefficient to construct a joint score, achieving risk constraint tree search. In replanning mode, only branches affected by events are locally reconstructed and rescored according to a smoothed and limited update frequency. After the optimal action is executed, the cost distribution prediction network, risk coefficient, and pruning margin are updated online in a closed loop based on the residual between the actual cost and the predicted cost, thereby achieving efficient, stable, robust, and continuously adaptive agent-based autonomous inference and decision-making.

[0010] To achieve the aforementioned objective, this invention provides, in one aspect, a method for constructing an agent's autonomous inference and decision-making model based on a decision tree, comprising:

[0011] A state vector is constructed based on environmental topology, observation data, and task constraints. Based on the state vector, the next state after each action in the action space is predicted, and an action feasibility indicator function is constructed based on the constraint margin of the next state. A set of rule-feasible candidate actions is obtained from the action space, and a set of supplementary candidate actions is generated based on the state vector. The union of the two is deduplicated to obtain a final set of candidate actions, and a decision tree is constructed based on the final set of candidate actions. A state-action input vector is constructed for each candidate action and input into a cost distribution prediction network to obtain the ensemble cost mean and prediction uncertainty for each candidate action. An event set is constructed based on environmental disturbances, and the event intensity is obtained from the event set. A risk coefficient is determined based on the event intensity. Based on the ensemble cost mean, the prediction uncertainty, the risk coefficient, and the final set of candidate actions, a decision tree is constructed. The action feasibility indicator function is used to construct a joint score, and risk-constrained pruning is performed on the decision tree based on the joint score, the current layer's optimal joint score, and the pruning margin. The replanning mode is switched according to the event intensity, mapping the event set to an event-affected subgraph, determining the event-affected branches, and performing local reconstruction and rescoring only on the event-affected branches according to an update frequency smoothed and limited based on the event intensity, to output the optimal action. After executing the optimal action, the prediction residual and normalized residual are calculated based on the actual cost of the action and the mean of the ensemble cost corresponding to the optimal action. When the prediction residual is greater than the residual threshold, the cost distribution prediction network is updated based on online samples, and the risk coefficient and pruning margin are updated synchronously according to the event intensity and the normalized residual.

[0012] To achieve the aforementioned objective, this invention also provides a system for constructing an autonomous inference and decision-making model for intelligent agents based on decision trees, comprising:

[0013] The system comprises several modules: a state modeling module for constructing state vectors based on environmental topology, observation data, and task constraints; a candidate action dual-source generation and decision tree structuring module for predicting the next state after each action in the action space based on the state vectors, constructing an action feasibility indicator function based on the constraint margin of the next state, filtering a set of rule-feasible candidate actions from the action space, generating a learning supplementary candidate action set based on the state vectors, removing duplicates from the union of the two sets to obtain a final candidate action set, and constructing a decision tree based on the candidate action set; a cost distribution prediction module for constructing state-action input vectors for each candidate action and inputting them into a cost distribution prediction network to obtain the integrated cost mean and prediction uncertainty for each candidate action; and an event intensity calculation and risk constraint tree search module for constructing an event set based on environmental disturbances, obtaining the event intensity based on the event set, determining the risk coefficient based on the event intensity, and calculating the risk coefficient based on the integrated cost mean and the predicted risk coefficient. A joint score is constructed using the measurement uncertainty, the risk coefficient, and the action feasibility indication function. Risk-constrained pruning is then performed on the decision tree based on the joint score, the current layer's optimal joint score, and the pruning margin. A local reconstruction and frequency control module for event-affected branches is used to control the switching of the replanning mode based on the event intensity. This module maps the event set to an event-affected subgraph, determines the event-affected branches, and performs local reconstruction and rescoring only on the event-affected branches in the replanning mode at an update frequency determined by a smoothing limit based on the event intensity, to output the optimal action. A residual closed-loop correction module is used to calculate the predicted residual and normalized residual based on the actual cost of the action and the mean of the integrated cost corresponding to the optimal action after executing the optimal action. When the predicted residual is greater than a residual threshold, the cost distribution prediction network is updated based on online samples, and the risk coefficient and pruning margin are updated synchronously based on the event intensity and the normalized residual.

[0014] Compared with the prior art, the present invention has at least the following beneficial effects:

[0015] 1. Balancing the constraints of feasibility and search coverage of candidate actions. This invention employs a dual-source candidate action generation method that combines a set of rule-based feasible candidate actions with a set of learned supplementary candidate actions. On the one hand, rule-based screening ensures that candidate actions meet constraints such as safe distance, time window, capacity, and resource margin. On the other hand, the learning supplementation mechanism expands the coverage of candidate actions, thereby reducing the omission of invalid actions and potential preferred actions, improving the quality of candidate actions and the effectiveness of subsequent searches.

[0016] 2. Improve the coupling between cost assessment and tree search. This invention uses a cost distribution prediction network to simultaneously output the integrated cost mean and prediction uncertainty of candidate actions, and constructs a joint score by combining action feasibility and risk coefficient. This allows the decision tree search process to no longer rely solely on single-point costs for branch selection, but to comprehensively consider the expected cost of the action, prediction risk, and constraint satisfaction, thereby improving the rationality of branch retention and pruning.

[0017] 3. Enhance decision-making robustness in dynamic environments. When disturbances such as changes in obstacles, capacity, or tasks occur in the environment, this invention can adaptively adjust the risk coefficient based on the intensity of the event, improving the tree search's sensitivity to candidate actions with high uncertainty. This reduces the probability of incorrectly retaining high-risk branches and enhances the system's decision-making stability and robustness under dynamic environments and complex task conditions.

[0018] 4. Reduce computational load for replanning and suppress jitter during mode switching. This invention does not perform a full reconstruction of the entire decision tree after an event occurs. Instead, it maps environmental events to subgraphs affected by the event, performing local reconstruction and rescoring only on the affected branches. Simultaneously, it controls replanning mode switching by incorporating entry thresholds, exit thresholds, minimum replanning intervals, and exit retention durations, and limits and smooths the update frequency, thereby reducing computational overhead, suppressing jitter caused by frequent replanning mode switching, and improving the system's real-time response capability.

[0019] 5. Establish an online closed-loop correction mechanism based on execution feedback. After the optimal action is executed, this invention constructs online samples based on the residual between the actual cost and the predicted cost of the action. When the predicted residual exceeds a threshold, the cost distribution prediction network is updated. Simultaneously, the risk coefficient and pruning margin are updated based on the event intensity and normalized residual. This allows the model prediction error and search parameter deviation to be corrected in a timely manner during the continuous decision-making process, thereby avoiding error accumulation and improving the long-term stability and adaptability of the system.

[0020] 6. The method and system structure are clearly correlated, facilitating engineering implementation. The implementation process of the method and the system module settings of this invention maintain strict correspondence in terms of state modeling, dual-source generation of candidate actions, cost distribution prediction, event intensity calculation and risk constraint tree search, local reconstruction and frequency control of branches affected by events, and residual closed-loop correction. The module boundaries are clear and the data flow is well-defined, which facilitates software deployment, modular implementation, and subsequent functional expansion. Attached Figure Description

[0021] Figure 1 This is a flowchart of the method for constructing an intelligent agent's autonomous inference and decision-making model based on decision trees provided by the present invention.

[0022] Figure 2This is a block diagram of the intelligent agent autonomous inference and decision model construction system based on decision tree provided by the present invention. Detailed Implementation

[0023] To enable those skilled in the art to more clearly understand the technical solutions of the present invention, the present invention will be further described in detail below with reference to embodiments. It should be noted that in this specific embodiment, the same letters represent the same meaning, and different letters represent different meanings; the same terms represent the same technical features, and different terms represent different technical features. This specific embodiment uses the autonomous passage and task execution scenario of a mobile intelligent agent in a dynamic architectural space as an example for illustration, but the present invention is not limited to this, and can also be applied to scenarios such as unmanned vehicle path planning, autonomous decision-making of inspection robots, warehouse handling scheduling, emergency evacuation simulation, and virtual task planning.

[0024] In this embodiment, the environmental topology is represented using a graph structure. Specifically, at time... The environmental topology diagram can be represented as: , in, For a moment Environmental topology diagram For a moment The set of nodes in the environmental topology graph. For a moment The set of edges in the environment topology graph. To avoid symbol misuse, in this specific implementation, it is always referred to as the edge set. The set of topological edges is represented by , and the set of events is denoted by . .

[0025] Figure 1 The flowchart illustrates the method for constructing an autonomous inference and decision-making model for an intelligent agent based on a decision tree, as provided in this invention. Figure 1 As shown, the method of the present invention includes steps S1 to S6. The state vector output by step S1 serves as the unified input for candidate action generation in step S2 and cost distribution prediction in step S3; the candidate action set and initial branch structure of the decision tree output by step S2 are input to step S4; the event intensity, joint score, and retained branch output by step S4 are input to step S5; the optimal action output by step S5 is executed, and the execution feedback is obtained by step S6 to form residuals, which are then used to update the cost distribution prediction network parameters in step S3 and the risk coefficient and pruning margin in step S4, thereby forming a complete technical closed loop around "state modeling - candidate action generation - risk constraint tree search - event-triggered local replanning - residual closed-loop correction".

[0026] Step S1: State Modeling

[0027] The input data for step S1 includes raw environmental information, agent observation data, and task constraint information; the output data for step S1 is a state vector. .

[0028] First, based on time Environmental topology diagram and time Node feature matrix Extraction time Topological coding The topology encoding can be represented as: , in, For topology-coded networks, These are the parameters of the topology-encoded network. Further, The node features include at least one or more of the following: node drivability, local congestion, obstacle occupancy status, and estimated travel time. In this embodiment, the topology coding network is preferably implemented using a graph neural network, so that topology information and node attribute information jointly participate in state representation learning.

[0029] Secondly, based on time Task constraint data construction time constraint margin vector The constraint margin vector can be expressed as:

[0030] , in, For a moment Safety distance margin For a moment Time window margin, For a moment The capacity margin, For a moment The resource surplus. Further, the resource surplus can correspond to energy surplus, computing power surplus, or communication resource surplus. Preferably, all of the above-mentioned constraint surpluses are characterized using a unified dimension or normalized form to facilitate the integrated processing of subsequent candidate action screening, cost prediction, and joint scoring.

[0031] Secondly, based on time Task instructions, target area, or target state construction time Target encoding The target encoding can be represented as: , in, For a moment The encoding result corresponding to the target location, target area, or target state. For a moment The encoding results correspond to the task priority, task stage, or task type. Preferably, the target encoding is represented in a vectorized manner so as to be concatenated and fused with the topology encoding and constraint margin vector.

[0032] Finally, the time Topological coding ,time constraint margin vector and time Target encoding By splicing the images together, we can obtain the time. state vector The state vector can be represented as: , in, This represents a vector concatenation operation. Thus, environmental topology information, constraint information, and target information are uniformly mapped to the state vector. Thus, the state vector output in step S1 is obtained. It can serve as the input for both the dual-source generation of candidate actions in step S2 and the cost distribution prediction in step S3, providing a unified data foundation for subsequent decision tree search.

[0033] In this step, the environmental topology and node attributes are first processed by a topology-coded network to form a topology code. The task constraint information is further converted into a constraint margin vector. The task objective information is further encoded into target code. The three are then concatenated to form a unified state vector. Therefore, this step achieves the coupling of "environmental information - constraint information - target information" into "unified state representation", which can reduce the candidate action selection bias, cost evaluation bias and tree search instability caused by scattered state definitions or inconsistent semantics.

[0034] This step unifies the coupling of topological structure information, constraint margin information, and task objective information into the state vector. In this process, the integrity, consistency, and computability of state representation can be improved, reducing candidate action selection bias, cost evaluation bias, and tree search instability caused by the dispersion of original data sources, inconsistent dimensions, or semantic inconsistencies. At the same time, it makes the generation of subsequent candidate actions more in line with the current environment and task constraints, and establishes cost prediction and tree search on a unified state semantic basis, thereby helping to improve the feasibility, accuracy, stability, and robustness of subsequent autonomous inference and decision-making processes.

[0035] Step S2: Dual-source generation of candidate actions and structuring of decision tree

[0036] The input data for step S2 includes the state vector output from step S1. and preset motion space The output data of step S2 includes a set of feasible candidate actions for the rules. Learn to supplement the candidate action set Candidate action set And the initial branch structure of the decision tree. The purpose of this step is to improve the coverage and exploratory nature of candidate actions while ensuring that the candidate actions meet the constraints of feasibility, and to organize the candidate actions into a unified decision tree structure, so as to provide action carriers and structural carriers for subsequent cost distribution prediction, joint score construction and risk constraint tree search.

[0037] First, based on time state vector Action space And the state transition model, for the action space Predict the next state after each action is performed to obtain the state at prediction time t+1 for each action. Predict the state at time t+1. It can be represented as: , in, For state transition model, The parameters of the state transition model are... For a moment The environmental dynamic disturbance parameters. Further, the state transition model preferably consists of kinematic update rules, resource consumption update rules, and constraint propagation rules. The kinematic update rules are used to predict changes in position, velocity, or orientation based on the current state and action; the resource consumption update rules are used to predict changes in resource reserves; and the constraint propagation rules are used to update safety distance reserves, time window reserves, and capacity reserves, thereby obtaining the state at prediction time t+1 corresponding to each action. .

[0038] Secondly, based on the state at prediction time t+1 corresponding to each action. Extract the constraint margin after performing the action and construct the state. Next action Feasibility indicator function The feasibility indication function for the action can be expressed as: , in, This is an indicator function; it takes the value 1 when the condition inside the parentheses is true, and 0 otherwise. To perform the action The safety distance margin at time t+1 later. To perform the action The remaining time window after time t+1, To perform the action The remaining capacity at time t+1. To perform the action The remaining resources at time t+1. This allows us to distinguish between actions that satisfy the constraints and those that do not, thus providing a basis for constructing the set of feasible candidate actions for the rules.

[0039] Furthermore, based on state Feasibility indicator function for each of the following actions From the action space Time obtained through filtering The set of feasible candidate actions according to the rules The set of feasible candidate actions for the rule can be represented as: , in, This establishes a preset action space. Through the above filtering, actions that meet the requirements of safe distance, time window, capacity, and resource margin can be prioritized, thereby expanding the set of feasible candidate actions for the rules. It has clear constraints and feasibility.

[0040] Secondly, based on time state vector Input the learning candidate generator network to obtain the action space. Recommended scores for each action, and select the highest-scoring actions. Each action as a moment Learning supplementary candidate action set The set of learning supplementary candidate actions can be represented as: , in, To learn the candidate generation network, The parameters of the learning candidate generation network are... To select the highest scorer Operators for each action, To preserve the number of actions, the learning candidate generation network is preferably obtained through offline supervised training or learning based on historical high-quality decision samples, so that the learning can supplement the candidate action set. It can supplement actions that are not covered by the rules but have potential value.

[0041] Then, the time The set of feasible candidate actions according to the rules With time Learning supplementary candidate action set Perform union and deduplication to obtain the time. Candidate action set The candidate action set can be represented as: , Therefore, on the one hand, rule-based selection ensures the feasibility of candidate actions under constraints, and on the other hand, learning and supplementation improve the coverage of candidate actions, thereby expanding the candidate action set. Balancing feasibility and exploratory aspects.

[0042] Finally, in terms of time state vector As the root node state, and based on time Candidate action set and the predicted next state for each candidate action Build Time Decision tree The decision tree can be represented as: , in, For a moment The set of decision tree nodes, For a moment The decision tree edge set. Furthermore, the root node of the decision tree carries the time... state vector Each tree edge corresponds to a candidate action, and each child node carries the state at the predicted time t+1 after executing the corresponding action. This provides a unified tree structure foundation for cost distribution prediction and risk constraint tree search in subsequent steps.

[0043] In this step, the state vector output by step S1 On the one hand, a set of feasible candidate actions is generated by the state transition model and feasibility judgment logic. On the other hand, a supplementary set of candidate actions is generated through a learning candidate generator network. The two are combined to form a candidate action set. And further organized into decision trees The initial branch structure is thus established. Therefore, step S2 achieves data coupling of "state representation - dual-source generation of candidate actions - action set fusion - decision tree structuring", which can provide a unified input basis for cost distribution prediction, joint score construction and local replanning in subsequent steps.

[0044] The beneficial effects of this step are as follows: by adopting a dual-source candidate action generation method of "rule-feasible candidate action set + learning-supplemented candidate action set", it can avoid insufficient action coverage caused by pure rule screening and reduce the risk of inactive actions generated by pure learning recommendation. This reduces the number of invalid branches at the source of the decision tree, improves the quality of candidate actions, and enhances the effectiveness, robustness and decision quality of subsequent tree search.

[0045] Step S3: Cost Distribution Prediction

[0046] The input data for step S3 includes the state vector output from step S1. and the candidate action set output by step S2 The output data of step S3 includes the average ensemble cost corresponding to each candidate action. and prediction uncertainty The purpose of this step is to map the state-action pair into a distributed cost evaluation result, so that subsequent steps can not only use the expected cost of candidate actions for decision tree search, but also use prediction uncertainty to suppress high-risk branches, thereby providing a unified quantitative basis for joint score construction and risk constraint tree search.

[0047] First, regarding time Candidate action set Each candidate action Based on time state vector and candidate actions Constructing the state-action input vector The state-action input vector can be represented as: , in, This represents the vector concatenation operation. Therefore, the time interval can be... The state information and candidate action information are uniformly mapped as the input to the cost distribution prediction network, so as to provide a unified input carrier for the subsequent calculation of the mean cost and prediction uncertainty.

[0048] Secondly, the state-action input vector Input cost distribution prediction network, for candidate actions Predicting the cost distribution yields the first... Individual networks at any time For candidate actions Mean of predicted cost of output The mean of the prediction cost It can be represented as: , in, For the first A cost distribution prediction subnetwork, For the first The parameters of the cost distribution predict the subnetwork. This refers to the sub-network index. Furthermore, the cost distribution prediction network is preferably implemented using a deep integration approach, that is, multiple sub-networks with identical structures but different parameters are set up to handle the same state-action input vector. To improve the stability and robustness of cost predictions, forecasts are made.

[0049] Then, based on the first Individual networks at any time For candidate actions Mean of predicted cost of output Integrate to obtain the time. Next candidate action Mean integration cost The average integration cost can be expressed as: , in, Predict the number of subnetworks based on the cost distribution. This represents the subnetwork index. By performing ensemble processing on the outputs of multiple subnetworks, the adverse effects of prediction bias from a single network on subsequent decision-making processes can be reduced.

[0050] Furthermore, based on the first Individual networks at any time For candidate actions Mean of predicted cost of output With the average integration cost The degree of dispersion between them determines the time. Next candidate action Prediction uncertainty The prediction uncertainty can be expressed as: , in, The number of cost distribution prediction subnetworks is used. Therefore, the degree of difference between multiple prediction results can be used to characterize the stability and reliability of the current candidate action cost prediction, thus providing a quantitative basis for the risk constraint term in subsequent joint scoring.

[0051] In another preferred embodiment, the cost distribution prediction network can also be implemented using a Monte Carlo random deactivation method, that is, keeping the randomly deactivated layers active during the inference phase and applying the same state-action input vector. Perform multiple forward predictions, and then obtain candidate actions based on the results of these multiple predictions. Mean integration cost and prediction uncertainty Therefore, cost distribution prediction can be achieved by using different uncertainty estimation methods without changing the overall process structure.

[0052] Furthermore, to ensure consistency in the cost definition used in the offline training and online execution phases, the time frame is adjusted based on the actual time consumption, actual risk cost, actual resource consumption, and constraint violation status after action execution. Actual cost of action Perform scalarization construction. The actual cost of the action can be expressed as: , in, For a moment The actual time cost, For a moment The actual risk and cost, For a moment The actual resource consumption cost, For a moment Penalties for violating the constraints , , and These are the weighting coefficients for the corresponding cost components. Through the above scalarization construction, the action execution results can be uniformly mapped to cost labels that can be used for training and correction, thus maintaining consistency with the aforementioned cost distribution prediction results.

[0053] Furthermore, the cost distribution prediction network is preferably constructed using a combination of offline pre-training and online incremental updates. In the offline stage, each sub-network can be trained separately based on historical state-action-cost samples to learn the mapping relationship between state-action pairs and cost distribution. In the online stage, the network parameters can be incrementally updated by combining execution feedback and residual results from subsequent steps, thereby improving the cost prediction network's adaptability to environmental changes, task changes, and distribution drift.

[0054] In this step, the state vector output by step S1 The candidate action set output by step S2 Each candidate action The concatenation forms a state-action input vector. The cost distribution prediction network is then used to obtain the ensemble cost mean for each candidate action. and prediction uncertainty Therefore, step S3 achieves data coupling of "state representation - candidate action - cost distribution prediction result", which can provide unified cost and risk inputs for joint score construction, risk constraint pruning and local replanning in subsequent steps.

[0055] The beneficial effects of this step are as follows: by adopting a cost distribution prediction method to jointly model the mean cost and prediction uncertainty of candidate actions, it can improve the adaptability of cost assessment results to environmental and task changes, and avoid the problems of high-risk branch retention or potential preferred branch pruning caused by relying solely on single-point cost estimation, thereby enhancing the robustness, stability and decision reliability of subsequent decision tree search.

[0056] Step S4: Event intensity calculation, joint score construction, and risk constraint tree search

[0057] The input data for step S4 includes the set of candidate actions output in step S2. and action feasibility indication function The average ensemble cost corresponding to each candidate action output in step S3. and prediction uncertainty and time Environmental disturbance events; the output data of step S4 includes time... event set Event intensity Risk coefficient The joint score corresponding to each candidate action Optimal joint score of the current layer And the set of retained branches. The purpose of this step is to couple the environmental disturbance, the cost assessment results of candidate actions, and the feasibility of actions into the same scoring framework, and to perform risk constraint screening on the candidate branches in the decision tree based on the scoring framework, so as to provide a unified quantitative basis for subsequent local reprogramming and optimal action search.

[0058] First, based on time Environmental disturbance event construction time event set The set of events can be represented as: , in, For a moment The One event, For a moment The number of events. Further, the events preferably include obstacle addition events, obstacle disappearance events, channel capacity change events, time window change events, and target position or target state change events. Thus, discrete disturbances in the environment can be uniformly organized into time-based events. event set This provides an input basis for subsequent event intensity calculations.

[0059] Secondly, based on time event set , No. The weight of each event and time The Quantization value of an event Calculate time Event intensity The intensity of the event can be expressed as: , in, For a moment The One event, For a moment The number of events, For the first The weight corresponding to each event For the first The quantified value of an event.

[0060] Furthermore, the event quantization function preferably employs different calculation methods based on the event type. For example, when the event is an obstacle addition event, the event quantization function can take the proportion of the newly added obstacle's coverage area to the area of ​​the locally passable area; when the event is a channel capacity change event, the event quantization function can take the channel capacity decrease ratio; when the event is a time window change event, the event quantization function can take the remaining time window reduction ratio; when the event is a target change event, the event quantization function can take the normalized value of the target position offset distance or the target state change amplitude. Thus, environmental disturbances of different types and dimensions can be uniformly mapped to time. Event intensity .

[0061] Furthermore, based on time Event intensity Determine the time risk coefficient The risk coefficient can be expressed as: , in, As the benchmark risk coefficient, Update the proportional coefficient to reflect the risk coefficient. This is the lower limit of the risk coefficient. This represents the upper limit of the risk coefficient. This is a limiting operator. Therefore, the risk coefficient can be reduced. With the intensity of the event The variable increases adaptively as the variable increases, thereby improving the sensitivity of the subsequent tree search process to prediction uncertainty when environmental disturbances increase.

[0062] Then, based on the time output in step S3 Next candidate action Mean integration cost Prediction uncertainty ,time risk coefficient and the status output in step S2 Next action Feasibility indicator function Construction time Next candidate action Joint rating The joint score can be expressed as: , in, This is the feasibility penalty coefficient. Therefore, the expected cost, predicted risk, and constraint feasibility of a candidate action can be uniformly mapped to a single score. This provides a unified criterion for subsequent branch retention and pruning.

[0063] Furthermore, based on time The following are candidate actions Joint rating Determine the time Optimal joint score of the current layer The optimal joint score of the current layer can be expressed as: , Therefore, the optimal scoring benchmark among the candidate actions in the current layer can be obtained, providing a reference baseline for subsequent risk constraint pruning.

[0064] Furthermore, based on time Optimal joint score of the current layer and time Pruning allowance The candidate branches in the decision tree are retained and pruned according to the branch retention criteria. The branch retention criteria can be expressed as: , Therefore, candidate branches whose joint scores are close to the current layer's best joint score can be retained, while candidate branches with significantly deteriorated scores can be removed, thereby reducing the computational load of subsequent searches while ensuring search quality.

[0065] In a preferred embodiment, the branch path is based on the multi-layered expansion. and the joint score of the corresponding actions at each layer along the path Construct branch paths Cumulative joint score The cumulative joint score for the path can be expressed as: , in, Extend the path depth This is the discount factor. For a moment Next path Layer corresponding actions The joint score is obtained. Thus, the local scores on multiple paths can be accumulated into a unified path-level evaluation value, so as to perform overall comparison and selection of candidate paths under multi-step deduction conditions.

[0066] In this step, environmental disturbance events are first organized into time intervals. event set And further quantified into moments. Event intensity ;time Event intensity Further decision-making time risk coefficient ;time risk coefficient Then, the average integration cost output from step S3 is used. Prediction uncertainty and the action feasibility indication function output in step S2 Jointly construct joint scoring Joint scoring Final joint score with the current layer's optimal score and pruning allowance Together, they determine which branches of the decision tree to retain and prune. Therefore, step S4 achieves data coupling of "event set - event intensity - risk coefficient - joint score - branch selection", which can provide a unified scoring basis for local tree reconstruction, optimal path search and optimal action output in subsequent steps.

[0067] The beneficial effects of this step are as follows: by incorporating event intensity, mean integration cost, prediction uncertainty, and action feasibility into the joint scoring construction process, it can avoid the problem of high-risk branch misretention caused by relying solely on single-point cost for branch selection, and improve the conservatism and stability of the tree search process when environmental disturbances increase or model uncertainty rises, thereby enhancing the rationality of branch retention, risk control capabilities, and the robustness and reliability of subsequent decision tree search.

[0068] Step S5: Local reconstruction of branches affected by events, hysteresis triggering, and update frequency control

[0069] The input data for step S5 includes the event set output in step S4. Event intensity Preserving branch sets and decision trees The output data of step S5 includes time. Subgraph affected by the event Replanning mode state, smoothed update frequency and optimal action The purpose of this step is to perform local reconstruction only on the decision tree branches that are truly affected by environmental events, and to stably drive the replanning process through hysteresis triggering and smooth control of update frequency, so as to reduce the computational jitter caused by invalid reconstruction and frequent switching while ensuring real-time responsiveness.

[0070] First, based on time event set Mapping environmental disturbances to time Subgraph affected by the event The event-affected subgraph can be represented as: , in, For a moment The set of affected nodes, For a moment The set of affected edges, This is a mapping function from an event to a local topological region. Further, when an event directly affects a node, the mapping function preferably returns the local neighboring nodes centered on that node and their associated edges; when an event directly affects an edge, the mapping function preferably returns that edge and its adjacent upstream nodes, downstream nodes, and a set of associated edges. Thus, the scope of an event's influence can be located from the global environment to a local topological region, thereby providing range constraints for subsequent local tree reconstruction.

[0071] Secondly, based on time Subgraph affected by the event Branch paths in decision trees Determine the branch path Whether it is a branch affected by the event. The condition for determining an affected branch can be expressed as: , in, branch path The resulting set of environment topology nodes and environment topology edges are obtained through mapping. For a moment The set of affected nodes, For a moment The set of affected edges, It is an empty set. Therefore, when the branch path Subgraph affected by the event When there is an intersection, the branch paths can be... Branches affected by the event are identified, and tree reconstruction and rescoring are performed only on the affected branches; for branches not affected by the event, their existing prediction state, joint score and cached results are kept unchanged, thereby avoiding a full reconstruction of the entire decision tree.

[0072] Furthermore, based on time Event intensity Entry threshold The last time we entered replanning mode and minimum replanning interval This determines whether to enter replanning mode. The conditions for entering replanning mode can be expressed as: ,

[0073] Therefore, only when time Event intensity The system only enters replanning mode when the entry threshold is reached and the time interval since the last entry into replanning mode is not less than the minimum replanning interval, thereby avoiding frequent replanning caused by short-term fluctuations in environmental events.

[0074] Furthermore, based on the exit threshold Duration of Exit Then, determine whether to exit the replanning mode. The condition for exiting the replanning mode can be expressed as: , in, For interval At any time within The intensity of the event, The exit threshold, To maintain the duration of exit and meet the following requirements. Therefore, the system exits the replanning mode only when the event intensity remains below the exit threshold for a continuous period of time not less than the exit duration. This hysteresis mechanism, which separates the entry and exit thresholds, suppresses frequent entry and exit from the replanning mode.

[0075] Then, in the replanning mode, based on time Event intensity Determine the time Pre-limit update frequency The update frequency before limiting can be expressed as: , in, To update the lower limit of the frequency, This is the update frequency scaling factor. Therefore, the update frequency can be adjusted according to the event intensity. The increase in the value enhances the system's ability to respond to environmental disturbances.

[0076] Furthermore, based on time Pre-limit update frequency Lower limit of update frequency and update frequency limit Determine the time Limiting update frequency The refresh rate after limiting can be expressed as: , in, This is a limiting operator. Therefore, it can prevent sudden increases or decreases in event intensity from causing the update frequency to exceed the system's tolerance range.

[0077] Furthermore, based on time Limiting update frequency and time Smoothed update frequency Perform first-order smoothing on the update frequency to obtain the time. Smoothed update frequency The smoothed update frequency can be expressed as: , in, This is a smoothing coefficient. Therefore, abrupt changes in update frequency between adjacent time points can be suppressed, thereby reducing computational jitter caused by frequency fluctuations during local tree reconstruction. In a preferred embodiment, when time... Event intensity When it increases, time Smoothed update frequency Monotonous and undiminished.

[0078] Furthermore, based on time Smoothed update frequency Determine the time Local reconstruction trigger interval The local reconstruction trigger interval can be expressed as: , Therefore, the smoothed update frequency can be This is further converted into a time-triggered rhythm for local reconstruction, so as to increase the frequency of local reconstruction when the event is intense and decrease the frequency of local reconstruction when the event is stable.

[0079] Finally, when each local reconstruction trigger interval is reached, the prediction time is recalculated only for the branches affected by the event. status Action Feasibility Indication Function and joint rating And perform a path search on the updated set of retained branches to obtain the time. Optimal path Furthermore, based on time... Optimal path Extract its first action to obtain the time. Optimal action The optimal action can be expressed as: , in, This is the operator for extracting the first action of the path. Therefore, while ensuring the local reconstruction range is controlled, the optimal action that adapts to changes in the current environmental events can be output. .

[0080] In this step, the set of events output by step S4 First, it is mapped to time. Subgraph affected by the event To limit the scope of local tree reconstruction; the event intensity output in step S4. Further decisions are made regarding whether the system enters or exits replanning mode, and the timing of these decisions is also determined. Pre-limit update frequency Update frequency after bandwidth limitation and smoothed update frequency ;time Smoothed update frequency This further determines the local reconstruction trigger interval. The branch affected by the event is recalculated and searched at each trigger interval, and the final output time is determined. Optimal action Therefore, step S5 achieves data coupling of "event localization - branch filtering - hysteresis triggering - frequency control - local reconstruction - optimal action output", which can provide a stable action output basis for subsequent execution feedback acquisition and residual closed-loop correction.

[0081] The benefits of this step are as follows: by performing local reconstruction only on branches affected by the event, a full update of the entire decision tree can be avoided, thereby reducing the computational load of replanning; by using a hysteresis triggering mechanism that separates entry and exit thresholds, frequent mode switching caused by environmental events fluctuating around the thresholds can be suppressed, thereby improving the stability of the replanning process; by performing amplitude limiting and smoothing processing on the update frequency, an adaptive balance can be achieved between real-time response capability and resource consumption, thereby enhancing the continuity, stability, and real-time performance of subsequent autonomous inference and decision-making processes.

[0082] Step S6: Residual closed-loop correction

[0083] The input data for step S6 includes the optimal action output in step S5. The optimal action output in step S3 The corresponding average integration cost and prediction uncertainty The event intensity output in step S4 and the time when the execution unit returns. Actual cost of action The output data of step S6 includes time. Predicted residuals Normalized residuals Updated cost distribution prediction network parameters Updated risk coefficient And the updated pruning allowance The purpose of this step is to perform online closed-loop correction on the cost distribution prediction network and risk constraint tree search parameters based on the real feedback after the optimal action is executed, so as to suppress the continuous accumulation of prediction errors and pruning biases, and provide updated model parameters and search parameters for the next round of autonomous inference and decision-making.

[0084] First, based on time Actual cost of action and optimal action The corresponding average integration cost Calculate time Predicted residuals The predicted residual can be expressed as: , in, This is an absolute value operation. Therefore, the optimal action can be determined. The deviation between the actual execution result and the predicted result is quantified as time interval. Predicted residuals This provides a basis for error analysis in subsequent online updates.

[0085] Secondly, based on time Predicted residuals and optimal action Corresponding prediction uncertainty Calculate time Normalized residuals The normalized residual can be expressed as: , in, To prevent zero constants, prediction errors can be uniformly normalized under different uncertainty levels, thereby improving the comparability of residual results across different actions, scenarios, and decision-making times.

[0086] Furthermore, based on time state vector Optimal action and the actual cost of the action Write the current execution feedback to the time. Online sample set The online sample set can be represented as: , in, For a moment Online sample set, This is a set union operation. Therefore, the actual execution feedback from the current decision-making process can be continuously accumulated into the online sample set, providing a data foundation for incremental updates of the subsequent cost distribution prediction network.

[0087] Furthermore, based on time Predicted residuals and preset residual threshold This determines whether to trigger an online update of the cost distribution prediction network. The online update triggering condition can be expressed as: , in, This is the residual threshold. Therefore, at time... Predicted residuals Exceeding the residual threshold If a significant deviation is found between the current cost prediction result and the actual execution result, the online update of the cost distribution prediction network can be triggered.

[0088] Then, based on time Actual cost of action Optimal action The corresponding average integration cost Prediction uncertainty and the cost distribution prediction network parameters after the last update. Construction time Online update loss The online update loss can be expressed as: , in, For stability regularity coefficients, Predict network parameters for the current cost distribution. It is a 2-norm, It is a natural logarithmic function. Therefore, while reducing the cost prediction error by utilizing the current online samples, the stability regularization term can suppress excessively large changes in network parameters, thereby improving the stability of the online update process.

[0089] Furthermore, based on time Online update loss Update the parameters of the cost distribution prediction network to obtain the time step. Updated cost distribution prediction network parameters The network parameter update process can be represented as follows: , in, For learning rate, Loss due to online updates Predict network parameters for the current cost distribution The gradient can be calculated. Therefore, the cost distribution prediction network can be updated online incrementally based on the latest execution feedback, thereby improving the adaptability of subsequent cost prediction results to environmental changes, task changes, and distribution drift.

[0090] Furthermore, based on time Event intensity Synchronous update time risk coefficient The risk coefficient update process can be expressed as: , in, As the benchmark risk coefficient, Update the proportional coefficient to reflect the risk coefficient. This is the lower limit of the risk coefficient. This represents the upper limit of the risk coefficient. This is a limiting operator. Therefore, the risk coefficient can be reduced. It adapts to changes in the level of environmental disturbance, thereby enabling the subsequent joint scoring to continuously match the current environmental risk status.

[0091] Furthermore, based on time Normalized residuals Synchronous update time Pruning allowance The pruning margin update process can be expressed as: , in, As a baseline pruning allowance, The pruning margin is updated to a new ratio coefficient. This is the lower limit of pruning allowance. This represents the upper limit of the pruning allowance. Therefore, the pruning allowance can be... It adapts to the level of prediction error, thereby dynamically adjusting the range of branches to be retained during subsequent tree search.

[0092] In this step, the optimal action output by step S5 is... The time is formed after execution Actual cost of action and the average integration cost output from step S3. and prediction uncertainty The time was calculated together Predicted residuals and normalized residuals The current execution feedback is further written into the online sample set. And when the online update trigger condition is met, it is used to update the parameters of the cost distribution prediction network. Meanwhile, the event intensity output in step S4 and time Normalized residuals Further used to update the risk coefficient and pruning allowance Therefore, step S6 achieves data coupling of "action execution feedback - residual calculation - online sample accumulation - model parameter update - search parameter update", which can provide an updated cost prediction model and risk constraint parameters for the next round of decision-making.

[0093] The beneficial effects of this step are as follows: by performing online closed-loop correction of the cost distribution prediction network, risk coefficient, and pruning margin based on real execution feedback, it can not only avoid the continuous accumulation of cost prediction errors and search parameter deviations in the continuous decision-making process, but also improve the model's adaptability to dynamic environments, task changes, and distribution drift, thereby enhancing the long-term stability, robustness, and decision reliability of the subsequent autonomous inference and decision-making process.

[0094] Figure 2 This is a block diagram of the intelligent agent autonomous inference and decision-making model construction system based on decision trees provided by this invention. Figure 2 As shown, this invention also provides a system for constructing an autonomous inference and decision-making model for an intelligent agent based on a decision tree. The system includes: a state modeling module, a dual-source candidate action generation and decision tree structuring module, a cost distribution prediction module, an event intensity calculation and risk constraint tree search module, a branch local reconstruction and frequency control module affected by events, and a residual closed-loop correction module. Each module can be implemented by a processor calling program instructions from memory, or by software, hardware, or a combination of both.

[0095] The state modeling module is used to construct a state vector based on environmental topology, observation data and task constraints, and send the state vector to the candidate action dual-source generation and decision tree structuring module and the cost distribution prediction module.

[0096] The dual-source generation and decision tree structuring module for candidate actions is used to predict the next state after each action is executed based on the state vector, action space, and state transition model, and to construct an action feasibility indicator function based on the constraint margin of the next state. It also filters out a set of rule-feasible candidate actions from the action space. Simultaneously, it generates a set of learning supplementary candidate actions based on the state vector, performs a union deduplication on the set of rule-feasible candidate actions and the set of learning supplementary candidate actions to obtain a set of candidate actions, and constructs a decision tree based on the set of candidate actions.

[0097] The cost distribution prediction module is used to construct a state-action input vector for each candidate action in the candidate action set, and input the state-action input vector into the cost distribution prediction network to obtain the integrated cost mean and prediction uncertainty corresponding to each candidate action, and send the integrated cost mean and prediction uncertainty to the event intensity calculation and risk constraint tree search module.

[0098] The event intensity calculation and risk constraint tree search module is used to construct an event set based on environmental disturbances, obtain the event intensity based on the event set, and determine the risk coefficient based on the event intensity; it is also used to construct a joint score based on the mean of the integrated cost, the prediction uncertainty, the risk coefficient and the action feasibility indication function corresponding to each candidate action, and perform risk constraint pruning on the decision tree based on the joint score, the current layer optimal joint score and pruning margin to obtain a set of retained branches.

[0099] The event-affected branch local reconstruction and frequency control module is used to control the switching of replanning mode according to the event intensity, map the event set into an event-affected subgraph, determine the event-affected branches in the decision tree, and perform local reconstruction and rescoring only on the event-affected branches in the replanning mode according to the update frequency after smoothing and limiting based on the event intensity, so as to output the optimal action.

[0100] The residual closed-loop correction module is used to calculate the predicted residual and normalized residual based on the actual cost of the action and the mean of the integrated cost corresponding to the optimal action after the optimal action is executed; when the predicted residual is greater than the residual threshold, the cost distribution prediction network is updated based on online samples, and the risk coefficient and the pruning margin are updated synchronously based on the event intensity and the normalized residual.

[0101] In this system, the state vector output by the state modeling module serves as the common input to the candidate action dual-source generation and decision tree structuring module and the cost distribution prediction module; the candidate action set, action feasibility indicator function, and decision tree output by the candidate action dual-source generation and decision tree structuring module serve as the input to the event intensity calculation and risk constraint tree search module; the integrated cost mean and prediction uncertainty output by the cost distribution prediction module serve as the input to the event intensity calculation and risk constraint tree search module; the event set, event intensity, and retained branch set output by the event intensity calculation and risk constraint tree search module serve as the input to the event-affected branch local reconstruction and frequency control module; the optimal action output by the event-affected branch local reconstruction and frequency control module serves as the input to the residual closed-loop correction module; the updated cost distribution prediction network, risk coefficient, and pruning margin of the residual closed-loop correction module are fed back to the corresponding modules, thus forming a closed-loop collaborative working process.

[0102] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit the scope of protection of the present invention. Without departing from the concept and essence of the present invention, those skilled in the art can still make various equivalent substitutions, modifications, or improvements to the technical solutions of the present invention. All technical solutions formed by adopting the technical features disclosed in the specification and claims of the present invention, or their equivalent technical features, should fall within the scope of protection of the present invention.

Claims

1. A method for constructing an agent's autonomous inference and decision-making model based on decision trees, characterized in that, include: A state vector is constructed based on environmental topology, observation data, and task constraints; Based on the state vector, predict the next state after each action in the action space is executed, and construct an action feasibility indicator function based on the constraint margin of the next state. Filter the action space to obtain a set of rule-feasible candidate actions, and simultaneously generate a set of learning supplementary candidate actions based on the state vector. Remove duplicates from the union of the two sets to obtain a final set of candidate actions, and construct a decision tree based on the set of candidate actions. For each candidate action, construct a state-action input vector and input it into a cost distribution prediction network to obtain the ensemble cost mean and prediction uncertainty corresponding to each candidate action. An event set is constructed based on environmental disturbances, and the event intensity is obtained based on the event set. A risk coefficient is determined based on the event intensity. A joint score is constructed based on the mean of the integrated cost, the prediction uncertainty, the risk coefficient, and the action feasibility indicator function. Risk constraint pruning is performed on the decision tree based on the joint score, the current layer's optimal joint score, and the pruning margin. The replanning mode is switched according to the event intensity. The event set is mapped to an event-affected subgraph, and the event-affected branches are determined. In the replanning mode, local reconstruction and rescoring are performed only on the event-affected branches according to the update frequency after smoothing and limiting based on the event intensity to output the optimal action. After executing the optimal action, the prediction residual and normalized residual are calculated based on the actual cost of the action and the mean of the integrated cost corresponding to the optimal action. When the prediction residual is greater than the residual threshold, the cost distribution prediction network is updated based on online samples, and the risk coefficient and the pruning margin are updated synchronously according to the event intensity and the normalized residual.

2. The method for constructing an agent-based autonomous inference and decision-making model based on decision trees as described in claim 1, characterized in that, The construction of the state vector based on environmental topology, observation data, and task constraints includes: based on time... Environmental topology diagram and time Node feature matrix Extraction time Topological coding The topological encoding can be represented as: ,in, For topology-coded networks, These are the parameters of the topology-coded network; Based on time Task constraint construction time constraint margin vector The constraint margin vector can be expressed as: ,in, For a moment Safety distance margin For a moment Time window margin, For a moment The capacity margin, For a moment The remaining resources; Based on time Task instructions, target area, or target state construction time Target encoding The target encoding can be represented as: ,in, For a moment The encoding result corresponding to the target location, target area, or target state. For a moment The coding results corresponding to task priority, task stage, or task type; Time Topological coding ,time constraint margin vector and time Target encoding By splicing the images together, we can obtain the time. state vector The state vector can be represented as: ， in, This indicates a vector concatenation operation.

3. The method for constructing an agent-based autonomous inference and decision-making model based on decision trees as described in claim 2, characterized in that, The process involves predicting the next state after each action in the action space based on the state vector, constructing an action feasibility indicator function based on the constraint margin of the next state, and filtering a set of rule-feasible candidate actions from the action space, including: based on time... state vector Action space And state transition model predicts action space Predicted time after each action is executed status Predicting the time status It can be represented as: ,in, For state transition model, The parameters of the state transition model are... For a moment Environmental dynamic disturbance parameters; Based on the predicted time of each action status Extract the constraint margin after performing the action and construct the state. Next action Feasibility indicator function The feasibility indication function of the action can be expressed as: ,in, For indicator functions, To perform the action Later time Safety distance margin To perform the action Later time Time window margin, To perform the action Later time The capacity margin, To perform the action Later time The remaining resources; Based on state Feasibility indicator function for each of the following actions From the action space Time obtained through filtering The set of feasible candidate actions according to the rules The set of feasible candidate actions for the rule can be represented as: 。 4. The method for constructing an agent-based autonomous inference and decision-making model based on decision trees as described in claim 3, characterized in that, The process of generating a learning supplementary candidate action set based on the state vector, removing duplicates from the union of the rule-feasible candidate action set and the learning supplementary candidate action set, and constructing a decision tree based on the candidate action set includes: based on time... state vector Input the learning candidate generator network to obtain the action space. Recommended scores for each action, and select the highest-scoring actions. Each action is a moment Learning supplementary candidate action set The set of learning supplementary candidate actions can be represented as: ,in, To learn the candidate generation network, The parameters of the learning candidate generation network are... To select the highest scorer Operators for each action, To preserve the number of actions; Time The set of feasible candidate actions according to the rules With time Learning supplementary candidate action set Perform union deduplication to obtain the time. Candidate action set The candidate action set can be represented as: ， In time state vector As the root node state, and based on time Candidate action set and the prediction time corresponding to each candidate action status Build Time Decision tree The decision tree can be represented as: ， in, For a moment The set of decision tree nodes, For a moment The decision tree edge set.

5. The method for constructing an agent-based autonomous inference and decision-making model based on decision trees as described in claim 4, characterized in that, The process of constructing a state-action input vector for each candidate action and inputting it into a cost distribution prediction network to obtain the ensemble cost mean for each candidate action includes: for time... Candidate action set Each candidate action Based on time state vector and candidate actions Constructing the state-action input vector The state-action input vector can be represented as: ， State-action input vector Input the cost distribution prediction network to obtain the first... Individual networks at any time For candidate actions Mean of predicted cost of output The mean of the prediction cost Represented as: ,in, For the first A cost distribution prediction subnetwork, For the first The parameters of the cost distribution predict the subnetwork. Subnet number; Based on the Individual networks at any time For candidate actions Mean of predicted cost of output Integrate to obtain the time. Next candidate action Mean integration cost The average integration cost can be expressed as: ,in, Predict the number of subnetworks based on the cost distribution. This is the sub-network sequence number.

6. The method for constructing an agent-based autonomous inference and decision-making model based on decision trees as described in claim 5, characterized in that, The process of obtaining the prediction uncertainty corresponding to each candidate action includes: based on the first... Individual networks at any time For candidate actions Mean of predicted cost of output With time Next candidate action Mean integration cost The degree of dispersion between them determines the time. Next candidate action Prediction uncertainty The prediction uncertainty can be expressed as: ， in, Predict the number of subnetworks based on the cost distribution. This is the sub-network sequence number.

7. The method for constructing an agent-based autonomous inference and decision-making model based on decision trees as described in claim 6, characterized in that, The process of constructing an event set based on environmental disturbances, obtaining event intensity based on the event set, and determining a risk coefficient based on the event intensity; and the process of constructing a joint score based on the mean of the integrated cost, the prediction uncertainty, the risk coefficient, and the action feasibility indicator function, and performing risk-constrained pruning on the decision tree based on the joint score, the current layer's optimal joint score, and pruning margin, includes: based on time... Environmental disturbance construction time event set The event set can be represented as: ； Based on time event set Weights corresponding to each event and event quantization function Calculate time Event intensity The intensity of the event can be expressed as: ， in, For a moment The One event, For a moment The number of events; Based on time Event intensity Determine the time risk coefficient The risk coefficient can be expressed as: ， in, As the benchmark risk coefficient, Update the proportional coefficient to reflect the risk coefficient. This is the lower limit of the risk coefficient. This represents the upper limit of the risk coefficient. For amplitude limiting operators; Based on time Next candidate action Mean integration cost Prediction uncertainty Risk coefficient and action feasibility indication function Construction time Next candidate action Joint rating The joint score can be expressed as: ,in, This is the feasibility penalty coefficient; Based on time The following are candidate actions Joint rating Determine the time Optimal joint score of the current layer The optimal joint score of the current layer can be expressed as: ； Based on time Optimal joint score of the current layer and time Pruning allowance The candidate branches in the decision tree are retained and pruned according to the branch retention condition, which can be expressed as: 。 8. The method for constructing an agent-based autonomous inference and decision-making model based on decision trees as described in claim 7, characterized in that, The step of controlling the replanning mode switching based on the event intensity, mapping the event set to an event-affected subgraph, and determining the event-affected branches includes: based on time... event set Mapping environmental disturbances to time Subgraph affected by the event The event-affected subgraph can be represented as: ， in, For a moment The set of affected nodes, For a moment The set of affected edges, This is a mapping function from events to local topological regions; Based on time Subgraph affected by the event Branch paths in decision trees Determine the branch path Whether a branch is affected by an event can be determined by the following criteria: ， in, branch path The resulting set of environment topology nodes and environment topology edges; It is an empty set; Based on time Event intensity Entry threshold The last time we entered replanning mode and minimum replanning interval To determine whether to enter replanning mode, the conditions for entering replanning mode can be expressed as: ； Based on exit threshold Duration of exit To determine whether to exit the replanning mode, the condition for exiting the replanning mode can be expressed as: ， in, For any moment within the time interval The intensity of the event, and .

9. The method for constructing an agent-based autonomous inference and decision-making model based on decision trees as described in claim 8, characterized in that, The reprogramming mode involves performing local reconstruction and rescoring only on the event-affected branches at an update frequency determined by a smoothing limit based on the event intensity, to output the optimal action; and after executing the optimal action, calculating the prediction residual and normalized residual based on the actual cost of the action and the mean ensemble cost corresponding to the optimal action, and updating the cost distribution prediction network based on online samples when the prediction residual is greater than a residual threshold, and synchronously updating the risk coefficient and pruning margin according to the event intensity and the normalized residual, including: in the reprogramming mode, based on time... Event intensity Determine the time Pre-limit update frequency The update frequency before limiting can be expressed as: ， in, To update the frequency ratio coefficient, This is the lower limit for the update frequency; Based on time Pre-limit update frequency Lower limit of update frequency and update frequency limit Determine the time Limiting update frequency The update frequency after the amplitude limiting can be expressed as: ； Based on time Limiting update frequency and time Smoothed update frequency Perform first-order smoothing on the update frequency to obtain the time. Smoothed update frequency The smoothed update frequency can be expressed as: ， in, For smoothing coefficients; Based on time Smoothed update frequency Determine the time Local reconstruction trigger interval The local reconstruction trigger interval can be expressed as: ； At the time of execution Optimal action Afterwards, based on time Actual cost of action and optimal action The corresponding average integration cost Calculation time Predicted residuals The predicted residual can be expressed as: ； Based on time Predicted residuals and optimal action Corresponding prediction uncertainty Calculation time Normalized residuals The normalized residual can be expressed as: ， in, To prevent zero constant; Based on time state vector Optimal action and the actual cost of the action Online sample sets The online sample set can be represented as: ， in, For a moment Online sample set; At that moment Predicted residuals Greater than the residual threshold At that time, predict network parameters based on the current cost distribution. Predicting network parameters based on cost distribution at time t-1 Construction time Online update loss The online update loss can be expressed as: ， in, For stability regularity coefficients, It is a 2-norm, It is the natural logarithm function; Based on time Online update loss Update the parameters of the cost distribution prediction network to obtain the time step. Updated cost distribution prediction network parameters The network parameter update process can be represented as follows: ， in, For learning rate, Loss due to online updates Predict network parameters for the current cost distribution gradient, Predict network parameters for the current cost distribution. Predict network parameters based on the cost distribution at time t-1; And based on time Event intensity Synchronous update time risk coefficient The risk coefficient update process can be expressed as: ， in, As the benchmark risk coefficient, Update the proportional coefficient to reflect the risk coefficient. This is the lower limit of the risk coefficient. This represents the upper limit of the risk coefficient. For amplitude limiting operators; And based on time Normalized residuals Update time Pruning allowance The pruning margin update process can be expressed as: ， in, As a baseline pruning allowance, The pruning margin is updated to a new ratio coefficient. This is the lower limit of pruning allowance. This represents the upper limit of pruning allowance.

10. A system for constructing an autonomous inference and decision-making model for an intelligent agent based on a decision tree, characterized in that, include: The state modeling module is used to construct state vectors based on environmental topology, observation data, and task constraints. The candidate action dual-source generation and decision tree structuring module is used to predict the next state after each action in the action space is executed based on the state vector, and construct an action feasibility indication function based on the constraint margin of the next state. It filters out a set of rule-feasible candidate actions from the action space, and simultaneously generates a learning supplementary candidate action set based on the state vector. The union of the two is deduplicated to obtain a candidate action set, and a decision tree is constructed based on the candidate action set. The cost distribution prediction module is used to construct a state-action input vector for each candidate action and input it into the cost distribution prediction network to obtain the mean integrated cost and prediction uncertainty corresponding to each candidate action. The event intensity calculation and risk constraint tree search module is used to construct an event set based on environmental disturbances, obtain the event intensity based on the event set, and determine the risk coefficient based on the event intensity. A joint score is constructed based on the mean of the integrated cost, the prediction uncertainty, the risk coefficient, and the action feasibility indicator function. Risk constraint pruning is then performed on the decision tree based on the joint score, the current layer's optimal joint score, and the pruning margin. The event-affected branch local reconstruction and frequency control module is used to control the switching of the replanning mode according to the event intensity, map the event set to an event-affected subgraph, determine the event-affected branches, and perform local reconstruction and rescoring only on the event-affected branches in the replanning mode according to the update frequency after smoothing and limiting based on the event intensity, so as to output the optimal action. The residual closed-loop correction module is used to calculate the predicted residual and normalized residual based on the actual cost of the action and the mean of the integrated cost corresponding to the optimal action after the optimal action is executed. When the predicted residual is greater than the residual threshold, the cost distribution prediction network is updated based on online samples, and the risk coefficient and the pruning margin are updated synchronously according to the event intensity and the normalized residual.

Images