Evaluation model optimization and dynamic calibration method based on meta-evaluation feedback

Abstract

The invention provides an evaluation model optimization and dynamic calibration method based on meta evaluation feedback, belongs to the field of artificial intelligence model evaluation, and constructs an evaluation system with a real-time feedback optimization capability through cooperative operation of a main evaluation model and a meta evaluation model in combination with a dynamic knowledge graph and a reinforcement learning framework. The main evaluation model is responsible for executing an original answer matching task, the meta evaluation model realizes error correction and logic optimization through variance analysis, semantic alignment and adversarial verification, the double models form a mutual verification mechanism through adversarial training, and dynamic disturbance is injected by utilizing a generative adversarial network to improve the robustness of the system. The adaptive ability of the evaluation system in a model iteration and knowledge extension scene is significantly improved, and a reliable evaluation criterion is provided for continuous optimization of an artificial intelligence model.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence model evaluation, in particular to an evaluation model optimization and dynamic calibration method based on meta-evaluation feedback. BACKGROUND

[0002] The traditional model evaluation system has long been plagued by the dual dilemmas of rigid evaluation standards and insufficient self-adaptive ability. Especially when the evaluated model is iteratively upgraded or the related field knowledge base is dynamically expanded, the evaluation model has difficulty in dynamically adjusting the judgment standard, resulting in systematic deviation of the evaluation result from the true performance of the model. SUMMARY

[0003] To solve the above technical problems, the present application provides an evaluation model optimization and dynamic calibration method based on meta-evaluation feedback. By introducing a meta-evaluation model and a four-dimensional feedback mechanism, combining the real-time updating ability of the field knowledge of the dynamic knowledge graph and the parameter optimization loop constructed by the reinforcement learning strategy, the performance degradation problem caused by the lag of the evaluation system is solved.

[0004] The present application aims to solve the problem of performance degradation caused by the lag of the evaluation system when the evaluated model is iteratively upgraded or the related field knowledge base is dynamically expanded. By using a dual-model collaborative architecture and a four-dimensional feedback mechanism, combining the real-time updating ability of the field knowledge of the dynamic knowledge graph and the parameter optimization loop constructed by the reinforcement learning strategy, the present application provides a reliable dynamic evaluation framework for the continuous iteration of artificial intelligence models.

[0005] The technical solution of the present application is:

[0006] An evaluation model optimization and dynamic calibration method based on meta-evaluation feedback, through the collaborative operation of the main evaluation model and the meta-evaluation model, combining the dynamic knowledge graph and the reinforcement learning framework, an evaluation system with real-time feedback optimization ability is constructed. The main evaluation model is responsible for performing the original answer matching task, while the meta-evaluation model realizes error correction and logic optimization through variance analysis, semantic alignment and adversarial verification. The dual model forms a mutual verification mechanism through adversarial training, and uses a generative adversarial network to inject dynamic disturbance to improve the system robustness.

[0007] Further,

[0008] Collaboration between the main evaluation model and the meta-evaluation model

[0009] The main evaluation model performs the original answer matching task

[0010] The main evaluation model serves as the basic execution layer, responsible for directly processing input data and generating preliminary results;

[0011] The meta-evaluation model performs secondary verification on the main model judgment

[0012] The meta-evaluation model improves the reliability of the main model through multi-dimensional verification, including:

[0013] Error correction: identify the bias of the main model through variance analysis and Kendall's coefficient of concordance, and dynamically adjust the weight;

[0014] Feedback optimization: guide the main model to iterate the reasoning path through step-level feedback;

[0015] Improve system robustness through adversarial training

[0016] Adversarial training enhances the model's resistance to noise and abnormal input by simulating attack scenarios, including:

[0017] Dynamic disturbance generation: use generative adversarial network (GAN) or projection adversarial training to inject small disturbances into input data, forcing the main model to learn more stable feature representation;

[0018] Double model mutual verification mechanism: the main model and the meta-model are mutually antagonistic in training;

[0019] Robustness evaluation index: combine statistical tests to quantify the performance degradation of the model in the adversarial environment, and guide the adjustment of training strategies.

[0020] Further,

[0021] Adopt four-dimensional feedback mechanism and deep integration of dynamic knowledge graph.

[0022] Among them,

[0023] The four-dimensional feedback mechanism includes:

[0024] Feature enhancement mechanism enhances the convergence efficiency of the model through high-frequency feature weight enhancement and interference suppression;

[0025] Gradient penalty mechanism uses a reverse correction strategy to correct incorrect decisions;

[0026] Feature reconstruction mechanism targets to repair confusion areas in the encoding layer;

[0027] Dynamic threshold mechanism realizes threshold safety calibration based on statistical monitoring window.

[0028] Dynamic knowledge graph forms a closed-loop linkage with the evaluation system through the integrity constraint of 12,000+ answer nodes, industry feature weight adaptation, and timeliness marking, and generates synthetic data to strengthen the knowledge evolution adaptability.

[0029] Further,

[0030] The reinforcement learning framework uses the PPO algorithm with multi-modal input to realize real-time mutual feedback between the evaluation strategy and the knowledge graph.

[0031] Further,

[0032] Feature reinforcement mechanism:

[0033] When the main model correctly matches the answer and the meta-model confirms that its confidence is reasonable, the feature reinforcement process is triggered: extract the key features of the correct answer from the semantic encoding layer of the main model, compare them with the feature library of historical correct samples, and identify the high-frequency activated feature dimensions; increase the gradient update weight of the neurons corresponding to the high-frequency features to make them converge faster during training; reduce the weight of non-core features through random masking technology to reduce the interference of redundant information on decision-making;

[0034] Gradient penalty mechanism:

[0035] When the main model incorrectly rejects the correct answer, and the meta-model confirms the error through adversarial sample verification, calculate the penalty intensity according to the confidence of the main model when it incorrectly rejects, the higher the penalty intensity, the greater the parameter correction force; during the backpropagation of the main model, reverse the gradient direction of the neurons that led to the incorrect rejection; set the maximum amplitude of single gradient update to avoid excessive punishment that damages the stability of the model;

[0036] Feature reconstruction mechanism:

[0037] When the main model incorrectly accepts the incorrect answer, and the meta-model detects such errors through adversarial sample detection, perform cluster analysis on the semantic features of the incorrect accepted answer to locate the confusion area of the encoding layer of the main model; according to the number of error samples, select specific channels of the encoding layer in proportion for reconstruction; only unfreeze the neurons of the selected channels for fine-tuning, and keep other parameters frozen to ensure that the main structure of the knowledge base is not affected;

[0038] Dynamic threshold mechanism:

[0039] Based on the adversarial verification results of the meta-model, dynamically calibrate the judgment threshold of the main model, set N test samples as the window, and count the confidence difference between the main model and the meta-model; if the difference persists, adjust the main model threshold by a fixed step; set a threshold fluctuation range to prevent extreme adjustment; at the same time, introduce a delay mechanism, which takes effect only when the trend of the last 3 monitoring windows is consistent.

[0040] Furthermore,

[0041] Dynamic knowledge graph

[0042] Coupling of answer integrity graph and double-model collaborative architecture

[0043] In the original answer matching phase of the main evaluation model, 12,000+ answer nodes preset in the knowledge graph are used as the semantic integrity benchmark library; through graph embedding technology, the answers are mapped into vectors, and when calculating the similarity with the user's answers, graph structure similarity constraints are introduced; the meta-evaluation model learns the matching bias pattern of the main model through adversarial training, and combines the node coverage state of the integrity graph to check whether the main model misses key knowledge branches;

[0044] Dynamic adaptation of feature weight graph and four-dimensional feedback mechanism

[0045] Based on the industry feature weight graph, a field-sensitive feature enhancement mechanism is designed, and the associated feature weight graph is synchronized: if the evaluation results of the model in a specific field deviate from the weight distribution of the graph, the gradient penalty coefficient of the feature in that field is automatically increased, forcing the model parameters to converge to the weight distribution labeled by the graph;

[0046] Dynamic threshold mechanism optimization driven by timeliness spectrum

[0047] The day / week / month frequency markers of the timeliness spectrum are designed in conjunction with the dynamic threshold mechanism to generate a time decay function; when the main model outputs the results, the confidence threshold is dynamically adjusted in conjunction with the timeliness markers;

[0048] Graph evolution mechanism and full-link closed loop of model parameters

[0049] An evaluation-graph-training closed loop is constructed, and the feature reconstruction mechanism is based on the evolved graph, using graph neural network GNN to encode node weights and connection relationships to generate synthetic data enhanced by field features, and injects the adversarial training set to improve the model's ability to adapt to knowledge changes.

[0050] Further,

[0051] Reinforcement learning framework

[0052] Deep integration of reward function and evaluation system

[0053] The reward mechanism is deeply integrated through the adversarial verification process of the collaborative architecture of the double model, realizing multi-level reward determination; through the dual verification of adversarial verification and dynamic threshold, the accuracy of the reward signal and the rationality of the exploration direction are ensured;

[0054] PPO algorithm and model architecture collaborative implementation

[0055] The policy network adopts a multi-modal input design, splices the main model confidence, the meta-model verification feature, the four-dimensional feedback index and the knowledge graph embedding vector into a 512-dimensional feature, and makes a decision through a 3-layer MLP with residual connection, i.e. 512-256-128 nodes; the feature reinforcement gate mechanism is introduced at the 256 node layer, and the information flow intensity is dynamically adjusted according to the feature weight in the four-dimensional feedback; the exploration strategy adopts a dynamic epsilon-greedy algorithm, and the initial exploration rate of 15% is automatically adjusted by sliding window statistics of the effective exploration rate.

[0056] The beneficial effects of the present application are

[0057] The standard rigidity and performance lag problems of the traditional evaluation system are effectively solved, the evaluation weight is dynamically adjusted, and the standard system evolves independently with the expansion of domain knowledge.

[0058] In terms of performance lag, the evaluation strategy and model upgrade are kept in synchronization and optimization, and the stable performance output of the evaluation system is ensured in continuous iteration. BRIEF DESCRIPTION OF DRAWINGS

[0059] Figure 1 is a workflow schematic diagram of the present application. DETAILED DESCRIPTION

[0060] In order to make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0061] The present application proposes an evaluation model optimization system and dynamic calibration method based on meta-evaluation feedback, aiming to solve the problems of standard rigidity, insufficient self-adaptability and performance lag existing in the traditional model evaluation system. The technical scheme cooperates the main evaluation model and the meta-evaluation model, combines the dynamic knowledge graph and the reinforcement learning framework, and constructs an evaluation system with real-time feedback optimization capability. The main evaluation model is responsible for performing the original answer matching task, while the meta-evaluation model realizes error correction and logic optimization through variance analysis, semantic alignment and adversarial verification. The double models form a mutual verification mechanism through adversarial training, and use the generative adversarial network to inject dynamic disturbance to improve the system robustness.

[0062] The core innovation lies in the deep integration of four-dimensional feedback mechanism and dynamic knowledge graph: the feature enhancement mechanism enhances the convergence efficiency of the interference suppression optimization model through high-frequency feature weight enhancement; the gradient penalty mechanism uses a reverse correction strategy to correct incorrect decisions; the feature reconstruction mechanism directionally repairs the confusion area of the encoding layer; and the dynamic threshold mechanism realizes threshold safety calibration based on a statistical monitoring window. The dynamic knowledge graph forms a closed-loop linkage with the evaluation system through the integrity constraint of 12,000+ answer nodes, industry feature weight adaptation, and timeliness marking, and generates synthetic data to strengthen the knowledge evolution adaptability. The reinforcement learning framework uses the PPO algorithm of multi-modal input to realize the real-time mutual feedback of the evaluation strategy and the knowledge graph.

[0063] The technical solutions of the present application are as follows:

[0064] b) Dual-model collaborative architecture

[0065] (4) The main evaluation model performs the original answer matching task

[0066] The main evaluation model serves as the basic execution layer and is responsible for directly processing input data and generating preliminary results. For example, in a question and answer system, the main model realizes the screening of the original answer through semantic vector matching or keyword retrieval. Such models usually require high precision but may have the limitation of a single perspective.

[0067] (5) The meta-evaluation model performs secondary verification on the main model

[0068] The meta-evaluation model (Critique Model) improves the reliability of the main model through multi-dimensional verification, including:

[0069] Error correction: identify the bias of the main model through variance analysis and Kendall's coefficient of concordance, and dynamically adjust the weight;

[0070] Feedback optimization: guide the main model to iterate the reasoning path through step-by-step feedback, avoiding logical errors caused by the single model capability bottleneck.

[0071] (6) Improve system robustness through adversarial training

[0072] Adversarial training enhances the model's resistance to noise and abnormal input by simulating attack scenarios, including:

[0073] Dynamic disturbance generation: use generative adversarial networks (GAN) or projection adversarial training to inject small disturbances into input data, forcing the main model to learn more stable feature representations;

[0074] Dual-model mutual verification mechanism: the main model and the meta-model are mutually antagonistic in training, for example, the main model generates adversarial samples, and the meta-model evaluates their rationality, forming a closed-loop optimization;

[0075] Robustness evaluation index: quantify the performance degradation of the model in the adversarial environment, guide the adjustment of training strategy.

[0076] c) Four-dimensional feedback mechanism

[0077] ⑤Feature enhancement mechanism:

[0078] When the main model correctly matches the answer and the meta-model confirms that its confidence is reasonable, the feature enhancement process is triggered: extract the key features of the correct answer from the semantic encoding layer of the main model, compare them with the feature library of historical correct samples, and identify the high-frequency activated feature dimensions; increase the gradient update weight of the neurons connected to the high-frequency features to make them converge faster during training; reduce the weight of non-core features through random masking technology to reduce the interference of redundant information on decision-making.

[0079] ⑥Gradient penalty mechanism:

[0080] When the main model incorrectly rejects the correct answer, and the meta-model confirms the error through adversarial sample verification, calculate the penalty intensity according to the confidence of the main model when it incorrectly rejects, the higher the penalty intensity, the greater the parameter correction strength; during the backpropagation of the main model, reverse the gradient direction of the neurons that led to the incorrect rejection; set the maximum amplitude of single gradient update to avoid excessive punishment that damages the stability of the model.

[0081] ⑦Feature reconstruction mechanism:

[0082] When the main model incorrectly accepts the wrong answer, and the meta-model detects such errors through adversarial sample detection, perform cluster analysis on the semantic features of the incorrectly accepted answer to locate the confusion area of the encoding layer of the main model; according to the number of error samples, select specific channels of the encoding layer for reconstruction in proportion; only unfreeze the neurons of the selected channels for fine-tuning, and keep other parameters frozen to ensure that the main structure of the knowledge base is not affected.

[0083] ⑧Dynamic threshold mechanism:

[0084] Based on the adversarial verification results of the meta-model, dynamically calibrate the judgment threshold of the main model, with a window of 500 test samples, and statistically analyze the confidence difference between the main model and the meta-model; if the difference persists, the system adjusts the threshold of the main model by a fixed step; set the threshold fluctuation range to prevent extreme adjustment. At the same time, introduce the "delay effect" mechanism, which only takes effect when the trend of the last 3 monitoring windows is consistent.

[0085] d) Dynamic knowledge graph

[0086] ⑤Coupling of answer integrity graph and dual-model collaborative architecture

[0087] In the original answer matching phase of the main evaluation model, 12,000+ answer nodes preset in the knowledge graph are used as the semantic integrity benchmark library. Through graph embedding technology, the answers are mapped to vectors, and when calculating the similarity with the user's answers, graph structure similarity constraints are introduced. The meta-evaluation model learns the matching bias patterns of the main model through adversarial training, and combines the node coverage state of the integrity graph to check whether the main model misses key knowledge branches.

[0088] ⑥Dynamic adaptation of feature weight graph and four-dimensional feedback mechanism

[0089] Based on the industry feature weight graph, a field-sensitive feature enhancement mechanism is designed, and the feature weight graph is associated in synchronization: if the evaluation results of the model in a specific field deviate significantly from the weight distribution of the graph, the gradient penalty coefficient of the feature in that field is automatically increased, forcing the model parameters to converge to the weight distribution labeled by the graph.

[0090] ⑦Dynamic threshold mechanism optimization driven by timeliness spectrum

[0091] The timeliness spectrum is marked with day / week / month frequency, and a time decay function is designed in conjunction with the dynamic threshold mechanism. When the main model outputs the results, the confidence threshold is dynamically adjusted in conjunction with the timeliness marker.

[0092] ⑧Graph evolution mechanism and full-link closed loop of model parameters

[0093] An evaluation-graph-training closed loop is constructed, for example, after each evaluation cycle, according to the consistent determination results of the main / meta model, the node weights of the knowledge graph are automatically adjusted according to the rule "update frequency = evaluation period x 0.3". At the same time, the feature reconstruction mechanism encodes the node weights and connection relationships based on the evolved graph using graph neural networks (GNN), generates synthetic data enhanced by field features, and injects the adversarial training set to improve the model's ability to adapt to knowledge changes.

[0094] e) Reinforcement learning framework

[0095] ③Deep integration of reward function and evaluation system

[0096] The reward mechanism is deeply integrated through the adversarial verification process of the dual model collaborative architecture, realizing multi-level reward determination. Through the dual verification of adversarial verification and dynamic threshold, the accuracy of the reward signal and the rationality of the exploration direction are ensured.

[0097] ④PPO algorithm and model architecture collaborative implementation

[0098] The policy network adopts a multi-modal input design, and a main model confidence, a meta-model verification feature, a four-dimensional feedback index and a knowledge graph embedding vector are spliced into a 512-dimensional feature, and a decision is made through a 3-layer MLP (512-256-128 nodes) with a residual connection. A feature reinforcement gate mechanism is introduced at the 256 node layer, and the information flow intensity is dynamically adjusted according to the feature weight in the four-dimensional feedback. The exploration strategy adopts a dynamic epsilon-greedy algorithm, and the initial exploration rate of 15% is automatically adjusted by a sliding window (100 steps) to count the effective exploration rate. The architecture realizes the real-time mutual feedback of policy decision and feature weight, and guarantees the balance between exploration and utilization.

[0099] The application significantly improves the adaptive ability of the evaluation system in the model iteration and knowledge expansion scene, and provides a reliable evaluation benchmark for the continuous optimization of artificial intelligence models.

[0100] The above description is only a preferred embodiment of the application, which is used to illustrate the technical solutions of the application, and is not used to limit the protection scope of the application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the application shall be included in the protection scope of the application.

Claims

1. A method for optimizing and dynamically calibrating an evaluation model based on meta-evaluation feedback, characterized in that: Through the coordinated operation of the main evaluation model and the meta-evaluation model, combined with the dynamic knowledge graph and reinforcement learning framework, an evaluation system with real-time feedback optimization capabilities is constructed; the main evaluation model is responsible for performing the original answer matching task; the meta-evaluation model realizes error correction and logic optimization through variance analysis, semantic alignment and adversarial verification. The two models form a mutual verification mechanism through adversarial training, and use generative adversarial networks to inject dynamic disturbances to improve system robustness.

2. The method according to claim 1, characterized in that Collaboration between the main evaluation model and the meta-evaluation model ① The main evaluation model performs the original answer matching task The main evaluation model serves as the basic execution layer, responsible for directly processing input data and generating preliminary results; ②The meta-evaluation model performs secondary verification on the main model judgment The meta-evaluation model improves the reliability of the main model through multi-dimensional validation, including: Error correction: Identify the deviation of the main model through variance analysis and Kendall's coefficient of harmony method, and dynamically adjust the weights; Feedback optimization: guiding the iterative reasoning path of the main model through step-level feedback; ③Improve system robustness through adversarial training Adversarial training enhances the model's resistance to noise and abnormal inputs by simulating attack scenarios. This is achieved through: Dynamic perturbation generation: Using generative adversarial networks (GANs) or projection adversarial training to inject small perturbations into the input data, forcing the main model to learn more stable feature representations; Dual-model mutual verification mechanism: The main model and meta-model act as adversaries during training; Robustness evaluation indicators: Combined with statistical tests to quantify the performance degradation of the model in adversarial environments and guide the adjustment of training strategies.

3. The method according to claim 1, characterized in that Adopt the deep integration of four-dimensional feedback mechanism and dynamic knowledge graph.

4. The method according to claim 3, characterized in that The four-dimensional feedback mechanism includes: The feature enhancement mechanism optimizes the model convergence efficiency by enhancing the high-frequency feature weights and suppressing interference; The gradient penalty mechanism uses a reverse correction strategy to correct wrong decisions; The feature reconstruction mechanism repairs the confusion areas of the coding layer in a targeted manner; The dynamic threshold mechanism implements threshold security calibration based on statistical monitoring windows.

5. The method according to claim 3, characterized in that The dynamic knowledge graph forms a closed-loop linkage with the evaluation system through the integrity constraints of 12,000+ answer nodes, industry feature weight adaptation and timeliness marking, generating synthetic data to enhance the adaptability of knowledge evolution.

6. The method according to claim 1, characterized in that The reinforcement learning framework adopts the PPO algorithm with multimodal input to achieve real-time feedback between the evaluation strategy and the knowledge graph.

7. The method according to claim 4, characterized in that ① Feature enhancement mechanism: When the main model correctly matches the answer and the meta-model confirms that its confidence is reasonable, the feature enhancement process is triggered: the key features of the correct answer are extracted from the semantic encoding layer of the main model, compared with the feature library of historical correct samples, and the feature dimensions with high-frequency activation are identified; the gradient update weights of the neuron connections corresponding to high-frequency features are increased to enable faster convergence during training; random masking technology is used to reduce the weights of non-core features to reduce the interference of redundant information on decision-making; ② Gradient penalty mechanism: When the main model mistakenly rejects the correct answer, and the meta-model confirms the error through adversarial sample verification, the penalty intensity is calculated based on the confidence level of the main model when it falsely rejects the answer. The higher the penalty intensity, the stronger the parameter correction. During the backpropagation phase of the main model, the gradient direction is reversed for the neuron connection that caused the false rejection. The maximum amplitude of a single gradient update is set to avoid damaging the model stability due to excessive penalty. ③ Feature reconstruction mechanism: When the main model mistakenly accepts an incorrect answer and the meta-model detects such an error through adversarial examples, it performs cluster analysis on the semantic features of the incorrectly accepted answer to locate the confusion areas in the main model's encoding layer. Based on the number of incorrect examples, it proportionally selects specific channels in the encoding layer for reconstruction. Only the neurons in the selected channels are unfrozen for fine-tuning, while other parameters remain frozen to ensure that the main structure of the knowledge base is not affected. ④Dynamic threshold mechanism: Based on the adversarial verification results of the meta-model, the judgment threshold of the main model is dynamically calibrated. N test samples are set as windows, and the confidence difference between the main model and the meta-model is counted. If the difference persists, the main model threshold is adjusted according to a fixed step size. The threshold fluctuation range is set to prevent extreme adjustments. At the same time, a delayed effectiveness mechanism is introduced, which will only take effect when the trends of three consecutive monitoring windows are consistent.

8. The method according to claim 4, characterized in that Dynamic Knowledge Graph ① Coupling of the answer completeness graph and the dual-model collaborative architecture During the primary evaluation model's raw answer matching phase, the 12,000+ answer nodes pre-populated in the knowledge graph serve as a semantic integrity benchmark. Graph embedding technology is used to map answers to vectors, and when calculating similarity with user answers, graph structure similarity constraints are introduced. The meta-evaluation model uses adversarial training to learn the matching deviation patterns of the primary model and, combined with the node coverage status of the integrity graph, verifies whether the primary model has omitted key knowledge branches. ② Dynamic adaptation of feature weight map and four-dimensional feedback mechanism Based on the industry feature weight map, we design a domain-sensitive feature enhancement mechanism and synchronize the associated feature weight map. If the model's evaluation results in a specific field deviate from the weight distribution of the map, we automatically increase the gradient penalty coefficient of the field feature to force the model parameters to converge to the weight distribution marked in the map. ③Optimization of dynamic threshold mechanism driven by time-sensitive personality spectrum The daily / weekly / monthly frequency markers of the time-sensitive character spectrum are linked with the dynamic threshold mechanism to design the time decay function; when the main model outputs the results, the confidence threshold is dynamically adjusted in combination with the time-sensitive markers; ④ Full-link closed loop of graph evolution mechanism and model parameters An evaluation-graph-training closed loop is constructed. At the same time, the feature reconstruction mechanism is based on the evolved graph, and a graph neural network (GNN) is used to encode node weights and connection relationships, generate synthetic data with enhanced domain features, and inject adversarial training sets to improve the model's adaptability to knowledge changes.

9. The method according to claim 6, characterized in that a) Reinforcement Learning Framework ① Deep integration of reward function and evaluation system The reward mechanism is deeply integrated with the adversarial verification process of the dual-model collaborative architecture to achieve multi-level reward determination. Through dual verification of adversarial verification and dynamic threshold linkage, the accuracy of reward signals and the rationality of exploration directions are ensured. ② Collaborative implementation of PPO algorithm and model architecture The policy network adopts a multimodal input design, concatenating the main model confidence, meta-model verification features, four-dimensional feedback indicators and knowledge graph embedding vectors into 512-dimensional features, and makes decisions through a three-layer MLP with residual connections, i.e. 512-256-128 nodes; a feature enhancement gating mechanism is introduced at the 256-node layer to dynamically adjust the information flow intensity according to the feature weights in the four-dimensional feedback; the exploration strategy adopts a dynamic ε-greedy algorithm, and the initial 15% exploration rate is automatically adjusted through the sliding window statistics of the effective exploration rate.

Images