Error Repair and Joint Evaluation Methods for Intelligent Models in Autonomous Control Systems

By constructing a repair context and generating candidate patches using multi-dimensional local repair operators, and combining metamorphic relations and micro-consistency constraints, the problem of local fault repair in deep reinforcement learning policy networks in autonomous control systems is solved, achieving safe and controllable model repair and performance improvement.

CN123115697APending Publication Date: 2026-11-17NORTHWESTERN POLYTECHNICAL UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611563420.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-10
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In autonomous control systems, deep reinforcement learning policy networks can develop policy biases due to unseen state inputs, leading to system performance degradation. Existing repair methods are time-consuming, prone to disrupting the correct behavior of the model, lack safety verification, and are difficult to apply on a large scale.

Method used

A repair context is constructed, candidate patches are generated through multi-dimensional local repair operators, and regression tests are performed in a simulation environment. The optimal repair scheme is selected by combining metamorphic relationships and micro-consistency constraints.

Benefits of technology

It achieves low-cost and secure local fault repair, maintains the overall model structure and correct behavior, improves the model's success rate and reward stability, and reduces time and computing power costs.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123115697A_ABST
    Figure CN123115697A_ABST
Patent Text Reader

Abstract

This application discloses a method for error repair and joint evaluation of intelligent models for autonomous control systems. The method includes constructing a repair context containing information about the deep reinforcement learning policy network to be repaired; obtaining a highly suspicious core subset based on the repair budget, enumerating multi-dimensional local repair operators and their discrete parameter grids to generate candidate patches; placing the candidate patches into a simulation environment to obtain macroscopic performance indicators, and constructing a metamorphic relationship conforming to control laws for microscopic consistency constraints; based on the macroscopic performance indicators and microscopic consistency constraints, selecting the optimal repair scheme and applying it to the deep reinforcement learning policy network. This application can reduce the governance time and computational cost of organizational control models, achieving safe and controllable model repair.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and data processing technology, specifically to a method for error repair and joint evaluation of intelligent models for autonomous control systems. Background Technology

[0002] With the rapid development of artificial intelligence technology, the application of autonomous control systems, represented by intelligent connected vehicles and drone swarms, in transportation, logistics, national defense, and industrial production has grown exponentially. In these systems, traditional rule-based and finite state machine-based control systems mainly follow an engineering paradigm of "finding patterns—modeling—decision-making." While this paradigm offers strong interpretability and formal verification capabilities, it often faces numerous challenges when dealing with highly dynamic, multi-agent game-like, and complex physical scenarios, including difficulties in mathematical modeling, insufficient rule coverage for edge scenarios, and high costs associated with scenario migration.

[0003] To overcome the limitations of traditional rule-based control, modern autonomous control systems increasingly rely on data-driven deep learning and reinforcement learning techniques. Deep reinforcement learning, through continuous closed-loop interaction between the agent and the environment, learns end-to-end policy mappings within the theoretical framework of Markov decision processes. Taking autonomous driving scenarios as an example, vehicles need to autonomously complete complex tasks such as lane keeping, automatic following, lane changing and overtaking, and collision avoidance control in traffic environments with multi-source sensor inputs, strong temporal constraints, and multi-participant interactions. Data-driven methods enable systems to develop powerful generalized decision-making capabilities without manually exhaustively enumerating all traffic rules, significantly reducing the cost of manual modeling. Typically, such systems employ deep Q-networks or policy gradient methods, using multi-layer nonlinear feature transformations to map the high-dimensional sensory state space to a continuous or discrete action value space, and relying on Bellman equations for iterative value updates.

[0004] However, while intelligent decision-making models bring leaps in system performance, they also introduce unprecedented engineering challenges. Because these models contain tens of thousands of neurons and high-dimensional connection weights, their decision-making mechanisms exhibit strong black-box characteristics. When deployed in real physical systems with long-tailed distributional biases, they are highly susceptible to policy biases due to unseen state inputs, potentially leading to serious safety incidents such as vehicle collisions and system malfunctions. Once the microscopic abnormal neurons causing macroscopic failures are accurately identified through a pre-emptive error localization mechanism, effectively repairing these faulty neurons becomes a key technical bottleneck hindering the large-scale and commercial application of autonomous control systems. Current engineering practice urgently needs a targeted, locally convergent, and safe regression verification method for automatically repairing models, considering the parameter space, structural space, and activation space of specific faulty neurons. Summary of the Invention

[0005] This application discloses a method, apparatus, electronic device, and storage medium for error repair and joint evaluation of intelligent models for autonomous control systems. It aims to solve the technical problem of how to automatically, precisely, safely, and cost-effectively repair the internal microscopic neuron defects that cause system performance degradation or decision-making errors in a label-free interactive environment lacking a complete testing language.

[0006] To achieve the above objectives, the first aspect of this application provides a method for error repair and joint evaluation of intelligent models for autonomous control systems, comprising: constructing a repair context, wherein the repair context includes the original parameters of the deep reinforcement learning policy network to be repaired, the simulation environment configuration, the historical interaction behavior trajectory dataset, and a set of suspicious neuron indices sorted in descending order of risk score; trunculating the set of suspicious neuron indices according to a preset repair budget to obtain a highly suspicious core subset; for the target neurons in the highly suspicious core subset, generating candidate patches containing a unified abstract representation by enumerating multi-dimensional local repair operators and their discrete parameter grids; placing each candidate patch into the simulation environment for regression testing to obtain macroscopic performance indicators, and constructing a metamorphic relationship that conforms to physical common sense to impose microscopic consistency constraints on the policy output; and based on the macroscopic performance indicators and the microscopic consistency constraint results, selecting the optimal repair scheme that meets the acceptance criteria and applying it to the deep reinforcement learning policy network.

[0007] Optionally, the construction of the repair context includes: encapsulating the original parameters, simulation environment configuration, historical interaction behavior trajectory dataset, suspicious neuron index set, and evaluation configuration parameter set required for joint verification into a five-tuple structure; wherein, the historical interaction behavior trajectory dataset includes state vectors, action distribution, and reward and punishment details.

[0008] Optionally, the multi-dimensional local repair operators include a weight perturbation operator, a weight regularization operator, a neuron pruning and local reconnection operator, and an activation amplitude pruning operator. Specifically, the weight perturbation operator injects independent and identically distributed zero-mean Gaussian noise into the local parameter neighborhood of the target neuron to break abnormal activation extrema; the weight regularization operator smooths the response of the target neuron through norm contraction to reduce its excessive dominance in global action value estimation; the neuron pruning and local reconnection operator performs logical pruning on ineffective neurons and uses feature cosine similarity to search for neighboring healthy neurons in the same layer, proportionally redistributing discarded outgoing edge weights to neighboring nodes to preserve the network's discriminative ability; the activation amplitude pruning operator inserts a dynamic truncation function during the forward inference stage of the computation graph, setting adaptive upper and lower bounds based on the statistical expectation and standard deviation of healthy interaction trajectories to isolate abnormal oversaturation signals.

[0009] Optionally, candidate patches containing a unified abstract representation are generated, including: for each target neuron in the highly suspicious core subset, the outer loop traverses all activated multi-dimensional local repair operator types, and the inner loop traverses the predefined discrete parameter grid for the operator; each time the inner loop hits, a patch instance containing a unique global identifier, operator type, target neuron topology index list and metadata information is generated through specific operator mapping logic, and it is added to the candidate patch queue.

[0010] Optionally, each candidate patch can be placed into a simulation environment for regression testing to obtain macro performance metrics, including: calculating the average undiscounted cumulative reward for all test rounds as the average return; calculating the variance of the cumulative reward fluctuation between different rounds as the return variance; the percentage of rounds in which the final goal is achieved according to the set environment clearance conditions statistical model as the absolute success rate; and monitoring the frequency of crash rounds that fall below the extremely low reward threshold as the extreme failure ratio.

[0011] Optionally, a transformation relationship conforming to physical common sense is constructed to impose micro-consistency constraints on the strategy output, including: performing physical noise robustness verification: given a basic safe traffic state, a small physical perturbation is applied to its sensor observation channel, and the optimal action category distribution output by the repaired model when processing the set of derived test pairs is detected to undergo a sudden reversal; performing spatial monotonicity verification: in the automatic following game scenario, a set of state sequences is constructed, and the longitudinal relative distance between the self-vehicle and the preceding vehicle is reduced in an orderly manner under the premise of strictly controlling the other variables to remain unchanged, and the value prediction for the acceleration breakthrough action shows a strictly monotonically decreasing trend, and the value prediction for the emergency braking action maintains a monotonically increasing trend.

[0012] Optionally, the optimal repair scheme that meets the acceptance criteria is selected, including: determining whether the metamorphosis relationship pass rate of the candidate patch is 100% and whether the degradation of the average return and success rate does not exceed the preset tolerance threshold; if it meets the criteria, it is determined to be technically acceptable; for all acceptable patches that have passed the acceptance, a comprehensive weighted scoring function is constructed, and a structural complexity penalty factor based on the total number of neurons and the number of stacking operators of a single patch is introduced for total order ranking, and the top patch is selected as the optimal repair scheme.

[0013] To achieve the above objectives, a second aspect of this application also provides an intelligent model error repair and joint evaluation device for autonomous control systems. The device comprises: a context construction module for constructing a repair context containing the original parameters of the deep reinforcement learning policy network to be repaired, simulation environment configuration, historical interaction behavior trajectory dataset, and a set of suspicious neuron indices sorted in descending order of risk score; a patch generation module for truncating the suspicious neuron index set according to a preset repair budget to obtain a highly suspicious core subset, and for the target neurons in the highly suspicious core subset, generating a candidate patch queue containing a unified abstract representation by enumerating multi-dimensional local repair operators and their discrete parameter grids; a joint verification module for placing each candidate patch into the simulation environment for regression testing to obtain macroscopic performance indicators, and constructing a metamorphic relationship consistent with physical common sense to constrain the policy output at the microscopic level; and a scheme output module for selecting the optimal repair scheme that meets the acceptance criteria based on the macroscopic performance indicators and the microscopic consistency constraint results, and applying it to the deep reinforcement learning policy network.

[0014] To achieve the above objectives, a third aspect of this application also provides an electronic device, including a processor, a memory, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the steps of the intelligent model error repair and joint evaluation method for autonomous control systems provided above.

[0015] To achieve the above objectives, the fourth aspect of this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method for error repair and joint evaluation of intelligent models for autonomous control systems provided above.

[0016] This application discloses a method for error repair and joint evaluation of intelligent models for autonomous control systems. The method includes constructing a repair context, which includes the original parameters of the deep reinforcement learning policy network to be repaired, the simulation environment configuration, the historical interaction behavior trajectory dataset, and a set of suspicious neuron indices sorted in descending order of risk score; trunculating the suspicious neuron index set according to a preset repair budget to obtain a highly suspicious core subset; for the target neurons in the highly suspicious core subset, generating candidate patches containing a unified abstract representation by enumerating multi-dimensional local repair operators and their discrete parameter grids; placing each candidate patch into the simulation environment for regression testing to obtain macroscopic performance indicators, and constructing a metamorphic relationship that conforms to physical common sense to impose microscopic consistency constraints on the policy output; based on the macroscopic performance indicators and microscopic consistency constraint results, selecting the optimal repair scheme that meets the acceptance criteria and applying it to the deep reinforcement learning policy network. Compared to existing deep learning debugging and model hardening technologies, this application brings significant and highly valuable engineering benefits to the field of autonomous control: First, it reduces time and computational costs: The traditional industry approach involves spending weeks collecting long-tail edge scene data that causes errors, and then using massive computing power to perform global fine-tuning and retraining on a model with tens of millions of parameters. The repair method in this application directly acts on microscopic neuron nodes, and the entire process of generating and verifying hot patches can converge in just a few minutes on conventional computing nodes. Second, it achieves safe and controllable model repair: Based on a joint verification strategy of multi-type neuron-level repair operators and metamorphic regression, local faults are repaired in a targeted manner without destroying the overall structure of the model and its original correct behavior (100% success rate). This avoids the high cost of overall retraining and ensures the safety of repair through constraints such as noise robustness and monotonicity, while improving the model's success rate and reward stability. Attached Figure Description

[0017] Figure 1 This is the overall flowchart of the intelligent model error repair method for autonomous control systems.

[0018] The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0019] It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit this application.

[0020] While academia and industry have explored error correction and security hardening of neural networks and deep reinforcement learning models, existing technologies still exhibit significant drawbacks and limitations when dealing with complex, interactive autonomous control systems. A detailed analysis of these shortcomings can be summarized along the following dimensions: First, global retraining is costly and prone to introducing regression defects. The industry's current common approach to remediation involves investing heavily in recollecting the scenario data that caused the error and adding it to the training set for global retraining or fine-tuning of the model. This method not only lacks targeting the root cause of the error, but more seriously, the global weight updates of deep reinforcement learning models often have a strong coupling effect. While eliminating a specific erroneous state, it easily damages the model's ability to represent other massive amounts of normal states, leading to a degradation of previously safe driving behavior. This remediation method, lacking unlabeled safety constraints, struggles to meet the high reliability requirements of industrial applications.

[0021] Second, provable repair methods based on formal verification suffer from extremely poor scalability. Some existing state-of-the-art research attempts to formulate the repair problem as a formal specification satisfying conditional constraints while maintaining locality and minimization. Upon discovering a violation input, a patch network is synthesized in the linear region where the violation occurs using a constraint solver, and then combined with the original network to correct the output. However, these methods typically rely heavily on piecewise linear activation structures and formally expressible logical specifications. When facing autonomous control tasks such as multi-source continuous state spaces, large parameter dimensions, and complex reward functions, such as those for autonomous driving, the computational overhead of formal counterexample search and verification explodes exponentially, making it impossible to scale to industrial-grade deep reinforcement learning policy networks.

[0022] Third, there is a lack of joint verification mechanisms to address the lack of test predictions. The verification of traditional software patch fixes heavily relies on a defined "input-output" labeled test set. However, in reinforcement learning, the environment state space is continuous and vast, involving multi-agent games, making it impossible for developers to predefine the absolutely correct optimal action for every traffic state. Existing tentative patching methods often rely solely on rerunning the simulation environment and observing changes in the "average reward" to evaluate the patch. However, this macroscopic scalar of average reward easily masks occasional fatal errors in long-term trajectories, lacking microscopic consistency constraints on patch fixes in common-sense dimensions such as "noise robustness" and "physical monotonicity," resulting in a lack of safety guarantees for the generated patch in real-world deployments.

[0023] In other words, in unlabeled interactive environments lacking comprehensive testing languages, when deep reinforcement learning policy networks are found to have internal micro-neural defects that lead to system performance degradation or decision-making errors, traditional remediation methods often rely on the recollection of global data and the overall retraining of the model. This crude approach not only consumes enormous computational and time resources but is also highly susceptible to major incidents. While fixing known defects, it can disrupt the correct decision-making logic and safety constraints already mastered by the model, leading to severe system regression risks. Furthermore, in autonomous control tasks, single-step actions are often difficult to directly classify as absolutely "correct" or "incorrect," and system success depends on long-term cumulative feedback. This characteristic makes it difficult to quickly validate the repaired model using traditional binary-labeled test cases. Existing tentative remediation methods often rely solely on rerunning the simulation environment and observing changes in the "average reward" to evaluate patches, lacking micro-consistency constraints on the patch's "noise robustness" and "physical monotonicity," resulting in a lack of security guarantees for the generated patch in real-world deployments.

[0024] To address a series of technical pain points in existing technologies, such as high global retraining costs, limited scalability of formal repair, inadequate runtime intervention, and lack of label-free security verification, this invention proposes a controllable and verifiable local error repair and joint evaluation method for autonomous control intelligent models. The core idea of ​​this technical solution is to model the repair process as a constrained parameter search centered on high-risk suspicious neurons, followed by local patch generation and multi-dimensional consistency constraint verification. Its basic content encompasses the following three core technical stages.

[0025] First, a multi-dimensional, low-intrusion neuron-level local repair operator system is designed. For the set of highly suspicious neurons output by the pre-error localization stage, this scheme breaks through the limitations of single parameter updates. It constructs four types of refined repair operations from three dimensions: parameter space, structure space, and activation space. These include weight perturbation, weight regularization, neuron pruning and local reconnection, and activation amplitude pruning. These operators, designed with minimal intrusion, fundamentally weaken the excessive dominance of defective neurons in global action value estimation by fine-tuning local parameters or dynamically constraining abnormal activation dynamics, without disrupting the overall network topology and normal feature extraction capabilities.

[0026] Secondly, a suspicion-driven local enumeration patch search and generation mechanism is constructed. To avoid computational overload caused by blind searching in the vast network parameter space, this method introduces the concept of repair context. Using neuron suspicion scores as a heuristic driving signal, a strict repair budget (such as the maximum neuron intervention ratio) is set for target truncation. On the locked target neurons, repair operators are automatically enumerated based on a pre-defined discrete parameter grid, generating a queue of candidate repair patches with a unified abstract representation. This strategy effectively reduces the global high-dimensional optimization problem to a combinatorial enumeration problem of a local candidate space.

[0027] Finally, a joint verification and patch acceptance criterion based on metamorphic relations and regression testing is proposed. To overcome the theoretical bottleneck of lacking test predictions, this scheme innovatively integrates metamorphic testing with reinforcement learning environment replay. For each generated candidate patch model, on the one hand, macroscopic performance indicators such as average reward, success rate, and extreme failure ratio are collected through environment interactive replay to evaluate the patch's repair benefits and stability; on the other hand, a metamorphic relation conforming to physical common sense is constructed to impose microscopic constraints on the logical consistency of policy outputs in the source state and perturbation state. Finally, the system performs a total order ranking of patches and outputs the optimal repair scheme with statistical safety guarantees based on a comprehensive scoring function and complexity penalty model that strictly adheres to "no performance degradation, maintained success rate, and satisfaction of metamorphic constraints".

[0028] To achieve the above basic objectives, the technical solution of this invention relies on a well-structured and logically rigorous automated workflow in its implementation. This section will elaborate on the specific algorithmic logic of the core repair context construction, the mathematical modeling of the repair operator, the local search strategy, and the joint verification mechanism within this workflow.

[0029] refer to Figure 1 The first embodiment of this application provides a method for error repair and joint evaluation of intelligent models for autonomous control systems. This addresses the technical problem presented in the background art: how to automatically, precisely, safely, and cost-effectively repair, at the parameter or structural level, faulty neurons in a deep reinforcement learning policy network when internal microscopic neuronal defects leading to system performance degradation or decision-making errors are detected, in an unlabeled interactive environment lacking a complete testing language. This method can be executed by a processor, which can be located on a terminal or server. The execution process of the method for error repair and joint evaluation of intelligent models for autonomous control systems is as follows: Step S101: Construct a repair context, which includes the original parameters of the deep reinforcement learning policy network to be repaired, the simulation environment configuration, the historical interaction behavior trajectory dataset, and a set of suspicious neuron indices sorted in descending order of risk score.

[0030] In one embodiment of this application, the original parameters, simulation environment configuration, historical interaction behavior trajectory dataset, suspicious neuron index set, and evaluation configuration parameter set required for joint verification are encapsulated into a five-tuple structure. The historical interaction behavior trajectory dataset includes state vectors, action distributions, and reward / penalty details.

[0031] Before implementing neuron-level repair, this method first needs to construct a unified repair context at the algorithm execution level, denoted as... This context provides the necessary model structure, data foundation, and constraint metrics for the subsequent repair engine, formally defined as a quintuple structure: In this formula, The original parameters of the deep reinforcement learning policy network to be repaired typically include the weight matrices of each layer (such as fully connected layers or convolutional layers). With bias vector . The simulation environment configuration represents the location of the agent, including the physical engine for Markov transition probabilities and the specific rule settings for the reward function. This is a dataset of historical agent interaction behavior trajectories, containing a series of states, actions, rewards, and internal neuron activation records. This is a set of suspicious neuron indexes provided by the pre-error localization mechanism and strictly ordered in descending order of risk score. This is the set of evaluation configuration parameters required for the joint validation module, such as the number of rounds in the regression test, the maximum number of steps per round, and the tolerance thresholds for each evaluation metric.

[0032] To unify the organization of the distinct repair operations in the parameter space, structure space, and activation space, this method proposes a general abstract representation structure for repair patches. Within the internal logic of the tool execution, a single patch... Defined as: In this notation, A unique global identifier is assigned to this patch for tracking within a complex verification queue. The specific operator type for the mapping has been defined, and its value range is defined as follows: . A list of topological indexes for the set of neurons to which this patch is applied. A parameterized vector required for a specific operator, whose specific dimensions and numerical meanings change dynamically depending on the type of operator. Metadata information used to record the generation of the patch, such as a memory snapshot of the original weights and the expected suspicion level, is used for rapid state rollback in case of patch verification failure. The state includes the suspicion value used during generation, coverage information, estimated risk level, and the identifier of the original weight snapshot used for rollback. This mathematical representation essentially defines the original parameter space... To the repaired parameter space Nonlinear mapping transformation equation .

[0033] Among them, For parameter space, These are the new parameters obtained after applying patch p.

[0034] Step S102: Based on the preset repair budget, the suspicious neuron index set is truncated to obtain a highly suspicious core subset. For the target neurons in the highly suspicious core subset, candidate patches containing a unified abstract representation are generated by enumerating multi-dimensional local repair operators and their discrete parameter grids. In an optional embodiment of this application, the multi-dimensional local repair operator includes a weight perturbation operator, a weight regularization operator, a neuron pruning and local reconnection operator, and an activation amplitude pruning operator. Specifically, the weight perturbation operator injects independent and identically distributed zero-mean Gaussian noise into the local parameter neighborhood of the target neuron to break abnormal activation extrema; the weight regularization operator smooths the response of the target neuron through norm contraction to reduce its over-dominance on global action value estimation; the neuron pruning and local reconnection operator performs logical pruning on ineffective neurons and uses feature cosine similarity to search for neighboring healthy neurons in the same layer, proportionally redistributing discarded outgoing edge weights to neighboring nodes to preserve the network's discriminative ability; the activation amplitude pruning operator inserts a dynamic truncation function during the forward inference phase of the computation graph, setting adaptive upper and lower bounds based on the statistical expectation and standard deviation of healthy interaction trajectories to isolate abnormal oversaturation signals.

[0035] Specifically, for any nth element in the target set Hidden layer A highly suspicious neuron is denoted as a node. Let the received incoming edge weight vector be... The corresponding bias scalar is The forward linear accumulation output of this neuron is defined as... ,in This is the activation output vector of the previous layer. The final nonlinear activation output of this neuron is... ,in Typically, nonlinear activation functions such as ReLU or Tanh are used. To comprehensively cover various potential defect modes, this invention precisely designs the following four types of orthogonal neuron-level repair operators.

[0036] 1. Weighted perturbation operator The physical significance of this operator lies in its ability to explore better configurations within the local parameter neighborhood of neurons trapped in locally poor solutions or exhibiting overfitting to specific features, thereby breaking their original abnormal activation extrema. Its mathematical transformation mapping... Defined as: ; The specific microscopic calculation formula is expanded as follows: ; ; In the above formula, is a discrete scaling factor. This is achieved by deliberately injecting independent, identically distributed, zero-mean Gaussian noise. The introduction of the noise term is crucial, as it provides a gradient-independent exploratory capability, enabling fine-tuning of activation amplitudes while avoiding large-scale modifications that could disrupt the original local curvature characteristics of neurons.

[0037] 2. Weighted Regularization Operator When mislocalization results indicate that a neuron excessively dominates the global action value, causing the agent to perform disastrous actions such as suddenly jerking the steering wheel, the operator uses hard... Norm contraction smooths the response. Incorporating this process into the retraining phase is equivalent to injecting an additional regularization penalty term into the objective function. Where β > 0 is the regularization coefficient, and w′ is the weight vector of neuron n under the patch parameter θ′.

[0038] In direct applications requiring no retraining, this operator degenerates into a closed-form coefficient decay: ; Among them, contraction factor This method, by configuring multiple sets of candidate regularization strengths, forcibly deprives the target neuron of its absolute say in subsequent network layers, thereby reducing the transmission effect of its abnormal peaks on macro-decision.

[0039] 3. Neuron pruning and local reconnection operators For neurons that are completely destroyed, utterly ineffective, or exhibit extremely chaotic representations, the most direct approach is to perform logical pruning, instantly severing the erroneous transmission pathways. This involves introducing a binary mask variable. This modifies the forward computation process to However, in complex decision networks, simple hard pruning can easily lead to a sharp drop in the overall representation capacity of the model and the collapse of the feature space. Therefore, this method creatively incorporates a local reconnection mechanism into the pruning logic.

[0040] Setting the first The weight matrix of the hidden layer is The outgoing edge weight vector projected by the pruned neuron to all nodes in the next layer is denoted as... The system will automatically search for and compare the pruned nodes using feature cosine similarity during algorithm execution. Healthy neurons in the same layer with similar orthogonal activation space characteristics constitute a candidate set. Using a preset reconnection ratio, the weights of discarded outgoing edges are proportionally redistributed to neighboring nodes: ; This reconnection operation constitutes a sophisticated topological compensation mechanism. While removing the lesions, it utilizes the representational redundancy of adjacent healthy neurons to preserve the network's original ability to discriminate complex continuous state spaces to the greatest extent, thus avoiding the cascading collapse caused by pruning.

[0041] 4. Activate the amplitude clipping operator To suppress uncontrolled extreme activation peaks in the model when faced with outliers in long-tailed distributions, this operator employs a non-parametric intrusive strategy. It does not permanently alter any physical weights, but instead inserts a dynamic truncation function in real-time during the forward inference phase of the computation graph: ; Boundary parameters It was not arbitrarily assigned, but rather based on strict data-driven criteria. This was achieved by extracting and repairing a massive amount of successful interaction trajectories from the context. The system calculates the statistical expectation of the healthy activation distribution of the neuron based on the data. with standard deviation Subsequently, through dynamic multiplier factors... Set adaptive upper and lower bounds: This truncation mechanism ensures that neurons remain completely transparent (without triggering truncation) under the vast majority of normal trajectory distributions, and only hard-limits and isolates abnormal oversaturation signals that are about to cause system catastrophe.

[0042] Furthermore, for each target neuron in the highly suspicious core subset, the outer loop iterates through all activated multi-dimensional local repair operator types, and the inner loop iterates through the predefined discrete parameter grid for that operator. Each time the inner loop hits, a patch instance containing a unique global identifier, operator type, target neuron topology index list and metadata information is generated through specific operator mapping logic and added to the candidate patch queue.

[0043] Specifically, after establishing the underlying operators for the above four dimensions, the core engineering challenge in the repair process is how to efficiently find the optimal solution in the vast network parameter space while avoiding combinatorial explosion. This method proposes a restricted local enumeration strategy driven by a list of decreasing suspicion levels.

[0044] For a list of suspicious neurons input by mislocation The system first uses the global repair budget hyperparameter Top_M to perform forced target truncation, stripping away low-risk nodes in the long tail to obtain a highly suspicious core subset. . Targeting For each target neuron, the outer loop of the algorithm iterates through all activated operator types, while the inner loop iterates through a predefined grid of discrete parameters for that operator. Each successful hit in the inner loop generates an abstract patch instance through specific operator mapping logic and adds it to the global candidate patch queue. This generation strategy cleverly reduces the dimensionality of the high-dimensional, non-convex, continuous parameter optimization problem to a combinatorial verification problem with a local finite candidate grid, fundamentally ensuring that the algorithm's time complexity is controllable and that the generated candidate set has sufficient diversity.

[0045] Step S103: Place each candidate patch into the simulation environment for regression testing to obtain macroscopic performance indicators, and construct a metamorphic relationship that conforms to physical common sense to impose microscopic consistency constraints on the strategy output.

[0046] In one embodiment of this application, the average cumulative reward without discount for all test rounds is calculated as the average return; the variance of the cumulative reward fluctuation between different rounds is calculated as the return variance; the percentage of rounds in which the final goal is achieved according to the set environmental clearance conditions statistical model is used as the absolute success rate; and the frequency of crash rounds that fall below the extremely low reward threshold is monitored as the extreme failure ratio.

[0047] For example, perform physical noise robustness verification: given a basic safe traffic state, apply a small physical perturbation to its sensor observation channel, and detect whether the optimal action category distribution output by the repaired model undergoes a sudden reversal when processing the set of derived test pairs; perform spatial monotonicity verification: construct a set of state sequences in an automatic following game scenario, and orderly reduce the longitudinal relative distance between the self-vehicle and the preceding vehicle while strictly controlling the other variables to remain unchanged, and detect whether the value prediction for acceleration breakthrough action shows a strictly monotonically decreasing trend, and whether the value prediction for emergency braking action maintains a monotonically increasing trend.

[0048] Specifically, for the patch queue To precisely select the optimal solution that improves policy performance without introducing regression defects, this method must overcome the validation bottleneck of deep reinforcement learning, which suffers from a lack of absolute action labels. Therefore, this technical solution constructs a joint validation architecture that integrates macro-statistics and micro-logic.

[0049] 1. Statistical Measurement of Macroscopic Regression Performance For each instantiated repair model in the candidate queue Place it in a defined simulation environment and execute it. The system involves long-term interactions spanning a full episode. During this process, the system closely monitors and calculates the following four macroscopic vector metrics: Average return The average undiscounted cumulative reward across all test rounds is used to characterize the absolute optimization of the strategy.

[0050] Return variance : Calculate the variance of cumulative rewards across different rounds as a proxy metric for measuring the stability of the repaired strategy against environmental disturbances.

[0051] Absolute success rate Based on the set environmental conditions for clearing the level (such as completing the route without collisions), the statistical model calculates the percentage of rounds in which the final goal is achieved. This is the most intuitive manifestation of safety.

[0052] Extreme failure rate Monitor the frequency of crash rounds that fall below extremely low reward thresholds and assess whether patches introduce more fatal logical dead ends.

[0053] 2. Consistency Verification of Microscopic Metamorphic Relationships Relying solely on average returns often masks distortions in deeper logic. This approach introduces metamorphic testing theory from software engineering, utilizing the inherent physical laws and causal logic of reinforcement learning environments to construct a set of metamorphic relationships. Implement label-free constraints: Physical noise robustness degradation relation: given any basic safe traffic state Apply a small physical perturbation conforming to a known Gaussian distribution to its sensor observation channel. The repaired model, when processing this set of derived test pairs, should not exhibit abrupt reversals in its output optimal action class distribution, and the action value vector should remain unchanged. of Fluctuations must be strictly bounded.

[0054] Spatial Monotonic Transformation Relationship: In a typical automated car-following game scenario, a sequence of states is constructed. While strictly controlling other variables (such as the vehicle's speed and lane occupancy status), the longitudinal relative distance between the vehicle and the vehicle in front is artificially and systematically reduced (i.e., the simulated environmental danger level monotonically increases). A logically sound policy network is used to predict the value of the "acceleration breakthrough" action. It must exhibit a strict monotonically decreasing trend, while the value prediction of the "emergency braking" action... It must remain monotonically increasing. Through directed sampling on the environment state tree, the system will output a Boolean-type consistency result indicator variable. This effectively filters out substandard patches that exploit loopholes or even cheating to obtain high returns, but whose physical logic is completely flawed.

[0055] Step S104: Based on the macroscopic performance indicators and microscopic consistency constraints, the optimal repair scheme that meets the acceptance criteria is selected and applied to the deep reinforcement learning policy network.

[0056] In an optional embodiment of this application, it is determined whether the metamorphosis pass rate of the candidate patch is 100% and whether the degradation of the average return and success rate does not exceed the preset tolerance threshold; if the conditions are met, it is determined to be technically acceptable; for all acceptable patch sets that have passed the acceptance test, a comprehensive weighted scoring function is constructed, and a structural complexity penalty factor based on the total number of neurons and the number of stacking operators of a single patch is introduced for full order ranking, and the top patch is selected as the optimal repair solution.

[0057] Specifically, based on the data collected above, the system ultimately undergoes a cold, hard acceptance test. A patch is defined as "technically acceptable" only if it satisfies the following strict inequality constraint: ;

[0058] The formula, in mathematical terms, stipulates that patches must 100% adhere to the physical laws of metamorphosis, while the rate of degradation in average return and success rate must not exceed a tiny tolerance threshold. .

[0059] For all accepted and excellent patch sets To select the final, unique, and fixed solution, the system constructed a comprehensive weighted scoring function: ; Finally, to implement the principle of "minimal intervention" in software engineering and prevent excessive, large-scale modifications from creating untraceable technical debt for long-term system operation, this invention introduces a penalty factor. This is calculated by statistically analyzing the total number of neurons affected by a single patch. Number of stacked operators Define structural complexity The final ranking list is based on a formula. Once generated, the system automatically captures the top patch and implements a network update, thus completing the entire process of seamless repair of the deep decision model.

[0060] Compared to existing deep learning debugging and model hardening techniques, this invention brings significant and highly valuable engineering benefits to the field of autonomous control: First, it reduces time and computing power costs: The traditional industry practice is to spend weeks collecting long-tail edge scene data that causes errors, and then use massive computing power to perform global fine-tuning and retraining on models with tens of millions of parameters. The repair method of this invention directly acts on microscopic neuron nodes, and the entire process of generating and verifying hot patches can converge in just a few minutes on conventional computing nodes.

[0061] Second, achieve safe and controllable model repair: Based on the joint verification strategy of multi-type neuron-level repair operators and metamorphic regression, local faults are repaired in a targeted manner without destroying the overall structure of the model and the original correct behavior (the original success rate is 100%). This avoids the high cost of overall retraining and ensures the safety of repair through constraints such as noise robustness and monotonicity, while improving the model success rate and reward stability.

[0062] The following section details the implementation process of an end-to-end, completely unsupervised, human-intervention-based neural network error repair and joint verification in a specific autonomous driving multi-lane merging environment, thereby comprehensively demonstrating the close linkage between various mathematical modules. However, those skilled in the art should understand that the specific operating environment and configuration thresholds described are for illustrative purposes only and are not intended to limit the scope of the claims.

[0063] Implementation Step 1: Construct the system repair context and identify the set of extremely high-risk neurons. First, initialize and load the fault model file that has completed policy output anomaly capture. Then, by connecting to the pre-installed spectrum and reward decay error localization module, read a list of neurons sorted in strict descending order of comprehensive risk suspicion score. Simultaneously, a dictionary set of historical interaction behavior data, including state vectors, action distributions, and reward / penalty details, is retrieved. Based on this, the system formally instantiates the repair context container in memory. To prevent model oscillations caused by the repair process, the program performs mandatory low-risk long-tail node truncation based on the pre-injected complexity control parameter Top_M. The system precisely extracts the top M neuron physical indices from the network containing thousands of nodes as the targets for this repair.

[0064] Implementation Step Two: Automated instantiation of multi-dimensional operators and generation of candidate patch queues Once the targets are identified, high-concurrency automatic operator enumeration and patch generation are performed on these M target nodes based on the configured discrete parameter grid. The system establishes a standardized 5-tuple metadata record for each generated configuration. After a rapid enumeration and combination process that takes only seconds, a highly diverse abstract queue of independent candidate patches was successfully generated, encompassing multiple intervention types. .

[0065] Step 3: Performance Regression Testing After the candidate queue is generated, it enters the verification phase. The scheduler... The candidate patches are iteratively dequeued based on the original model parameters. Regression tests were performed one by one, with a fixed success threshold and random seed. The effectiveness of the patches was evaluated from dimensions such as average reward, success rate, original success retention rate, and action offset. Acceptable patches that met the criteria of "transformation relationship pass rate and no significant performance degradation" were selected. Finally, the optimal patch, model comparison report, and reproduction data file were output to solidify and trace the repair results.

[0066] The second embodiment of this application provides an autonomous control model error repair device to solve the same technical problem as the method embodiment. The device may include the following modules: a context construction module, a patch generation module, a joint verification module, and a scheme output module.

[0067] The system comprises the following modules: a context construction module, a repair context containing the original parameters of the deep reinforcement learning policy network to be repaired, simulation environment configuration, historical interaction behavior trajectory dataset, and a set of suspicious neuron indices sorted in descending order of risk score; a patch generation module, a high-suspicious core subset, which is obtained by truncating the suspicious neuron index set according to a preset repair budget, and a candidate patch queue containing a unified abstract representation is generated for the target neurons in the high-suspicious core subset by enumerating multi-dimensional local repair operators and their discrete parameter grids; a joint verification module, a microscopic consistency constraint, which is constructed to constrain the policy output based on the macroscopic performance index and the microscopic consistency constraint results; and a scheme output module, which selects the optimal repair scheme that meets the acceptance criteria based on the macroscopic performance index and the microscopic consistency constraint results and applies it to the deep reinforcement learning policy network.

[0068] It is understood that the above-described device embodiments correspond to the method embodiments. Therefore, the device embodiments have all the beneficial effects of the method embodiments, and will not be repeated here.

[0069] The third embodiment of this application provides an electronic device to solve the same technical problem as the method for error repair and joint evaluation of intelligent models for autonomous control systems provided in the previous embodiments. The electronic device may include a processor, a memory, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of the method for error repair and joint evaluation of intelligent models for autonomous control systems provided in the previous embodiments.

[0070] The fourth embodiment of this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method for error repair and joint evaluation of intelligent models for autonomous control systems provided in the preceding embodiments.

[0071] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for intelligent model error correction and joint evaluation of autonomous control systems, characterized in that, include: Construct a repair context, which includes the original parameters of the deep reinforcement learning policy network to be repaired, the simulation environment configuration, the historical interaction behavior trajectory dataset, and a set of suspicious neuron indices sorted in descending order of risk score; Based on the preset repair budget, the suspicious neuron index set is truncated to obtain a highly suspicious core subset. For the target neurons in the highly suspicious core subset, candidate patches containing a unified abstract representation are generated by enumerating multi-dimensional local repair operators and their discrete parameter grids. Each candidate patch is placed into the simulation environment for regression testing to obtain macroscopic performance indicators, and a metamorphic relationship that conforms to physical common sense is constructed to impose microscopic consistency constraints on the policy output. Based on the macroscopic performance indicators and microscopic consistency constraints, the optimal repair scheme that meets the acceptance criteria is selected and applied to the deep reinforcement learning policy network.

2. The method for error repair and joint evaluation of intelligent models for autonomous control systems as described in claim 1, characterized in that, The build repair context includes: The original parameters, simulation environment configuration, historical interaction behavior trajectory dataset, suspicious neuron index set, and evaluation configuration parameter set required for joint verification are encapsulated into a 5-tuple structure: , in, These are the original parameters of the deep reinforcement learning policy network to be repaired; The simulation environment configuration representing the intelligent agent; A dataset of agent interaction behavior trajectories collected historically; This is a set of suspicious neuron indices provided by the pre-error localization mechanism and strictly ordered in descending order of risk score; This is the set of evaluation configuration parameters required for the joint validation module, such as the number of rounds in the regression test, the maximum number of steps per round, and the tolerance thresholds for various evaluation metrics.

3. The method for error repair and joint evaluation of intelligent models for autonomous control systems as described in claim 1, characterized in that, The multidimensional local repair operators include weight perturbation operators, weight regularization operators, neuron pruning and local reconnection operators, and activation amplitude pruning operators; The weight perturbation operator injects independent and identically distributed zero-mean Gaussian noise into the local parameter neighborhood of the target neuron to break the abnormal activation extrema. The weight regularization operator smooths the response of the target neuron through norm contraction to reduce its over-dominance on the global action value estimate. The neuron pruning and local reconnection operators perform logical pruning on failed neurons and use feature cosine similarity to search for neighboring healthy neurons in the same layer. The discarded outgoing edge weights are proportionally redistributed to neighboring nodes to preserve the network's discriminative ability. The activation amplitude pruning operator inserts a dynamic truncation function during the forward inference phase of the computation graph, and sets adaptive upper and lower bounds based on the statistical expectation and standard deviation of the healthy interaction trajectory to isolate abnormal oversaturation signals.

4. The method for error repair and joint evaluation of intelligent models for autonomous control systems as described in claim 1, characterized in that, The generation of candidate patches containing a unified abstract representation includes: For each target neuron in the highly suspicious core subset, the outer loop iterates through all activated multi-dimensional local repair operator types, and the inner loop iterates through the predefined discrete parameter grid for that operator. Each time an inner loop hits, a patch instance containing a unique global identifier, operator type, target neuron topology index list, and metadata information is generated through specific operator mapping logic and appended to the candidate patch queue.

5. The method for error repair and joint evaluation of intelligent models for autonomous control systems as described in claim 1, characterized in that, The step of placing each candidate patch into the simulation environment for regression testing to obtain macroscopic performance indicators includes: The average return is calculated as the average cumulative reward without discounts across all test rounds. Calculate the variance of cumulative reward fluctuations across different rounds as the return variance; The absolute success rate is the percentage of rounds in which the final goal is achieved based on a statistical model of the established environmental conditions. The frequency of crash rounds that fall below the extremely low reward threshold is monitored as the proportion of extreme failures.

6. The method for error repair and joint evaluation of intelligent models for autonomous control systems as described in claim 1, characterized in that, The constructed metamorphic relationship, which conforms to the laws of physics, imposes microscopic consistency constraints on the strategy output, including: Perform physical noise robustness verification: Given a basic safe traffic state, apply a small physical perturbation to its sensor observation channel, and detect whether the optimal action category distribution output by the repaired model undergoes a sudden reversal when processing this set of derived test pairs; Execution of spatial monotonicity verification: In the automatic following game scenario, a set of state sequences is constructed. Under the premise of strictly controlling other variables to remain unchanged, the longitudinal relative distance between the self-vehicle and the preceding vehicle is reduced in an orderly manner. The value prediction for the acceleration breakthrough action shows a strictly monotonically decreasing trend, and the value prediction for the emergency braking action maintains a monotonically increasing trend.

7. The method for error repair and joint evaluation of intelligent models for autonomous control systems as described in claim 1, characterized in that, The selection of the optimal repair solution that meets the acceptance criteria includes: Determine if the metamorphosis success rate of candidate patches is 100% and the degradation of average return and success rate does not exceed the preset tolerance threshold; if so, it is considered technically acceptable. For all acceptable patches that have passed the acceptance test, a comprehensive weighted scoring function is constructed, and a structural complexity penalty factor based on the total number of neurons and the number of stacking operators based on the effect of a single patch is introduced for total order ranking. The top-ranked patch is then selected as the optimal repair solution.

8. A device for error repair and joint evaluation of intelligent models for autonomous control systems, characterized in that, include: The context building module is used to construct a repair context containing the original parameters of the deep reinforcement learning policy network to be repaired, the simulation environment configuration, the historical interaction behavior trajectory dataset, and a set of suspicious neuron indices sorted in descending order of risk score; The patch generation module is used to target truncate the suspicious neuron index set according to the preset repair budget to obtain a highly suspicious core subset. For the target neurons in the highly suspicious core subset, a candidate patch queue containing a unified abstract representation is generated by enumerating multi-dimensional local repair operators and their discrete parameter grids. The joint verification module is used to put each candidate patch into the simulation environment for regression testing to obtain macroscopic performance indicators, and to construct a metamorphic relationship that conforms to physical common sense to constrain the microscopic consistency of the policy output. The solution output module is used to select the optimal repair solution that meets the acceptance criteria based on the macro performance indicators and micro consistency constraint results, and apply it to the deep reinforcement learning policy network.

9. An electronic device, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the steps of the method for error repair and joint evaluation of intelligent models for autonomous control systems as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements the steps of the method for error repair and joint evaluation of intelligent models for autonomous control systems as described in any one of claims 1 to 7.