A communication agent meta-learning driven adaptive evolution and cross-domain collaboration method and system

Abstract

This invention belongs to the field of artificial intelligence communication technology and discloses a method and system for adaptive evolution and cross-domain collaboration driven by meta-learning of communication agents. Addressing the specific technical problems of communication agents, such as weak rapid adaptation to new scenarios, insufficient cross-domain collaborative knowledge transfer, and lack of autonomous optimization mechanisms for dynamic evolution, this invention constructs a meta-learning adaptive modeling framework for communication agents, designs cross-domain collaborative knowledge distillation and transfer algorithms, and proposes an autonomous optimization mechanism for the dynamic evolution of agents. This achieves a technical upgrade for communication agents from fixed model adaptation, isolated knowledge application, and passive iterative upgrades to rapid scenario adaptation, cross-domain knowledge reuse, and proactive evolution optimization.

Description

Technical Field

[0001] This invention belongs to the field of information technology, specifically relating to a method and system for adaptive evolution and cross-domain collaboration driven by meta-learning of a communication intelligent agent. Background Technology

[0002] As the core carrier of the intelligent evolution of communication networks, intelligent communication agents need to adapt to diverse scenarios (such as industrial internet, smart cities, and emergency communications), support cross-domain collaborative services, and achieve long-term dynamic evolution. They are widely used in scenarios such as 5G / 6G heterogeneous networks, edge cloud collaborative communication, and global communication management. Existing intelligent communication agent technologies are mostly based on training for specific scenarios and isolated functional designs, lacking deep algorithmic design for rapid adaptation to new scenarios, reusability of cross-domain knowledge, and autonomous self-evolution. In practical applications, there are three specific and urgent technical problems that need to be solved, as follows: The ability to quickly adapt to new scenarios is weak, and it relies on a large amount of labeled data and long-term training: The training of existing communication agents models mostly depends on a large amount of labeled data in specific scenarios. When faced with new scenarios (such as adding new edge node types or changes in business transmission protocols), it is necessary to re-collect data and perform full training. The adaptation cycle is long and costly, and it is impossible to utilize the adaptation experience of historical scenarios, resulting in low deployment efficiency for new scenarios.

[0003] Insufficient knowledge transfer in cross-domain collaboration leads to knowledge barriers and repetitive learning: The knowledge system of existing communication agents is limited to their own deployment domain. When collaborating across domains, they can only interact based on surface-level instructions and cannot effectively transfer core knowledge (such as scenario adaptation strategies and resource scheduling experience). Agents from different domains need to repeat learning for similar problems, resulting in low efficiency and poor knowledge reuse in cross-domain collaboration.

[0004] Dynamic evolution lacks an autonomous optimization mechanism and relies on manual iterative upgrades: The model optimization and functional upgrades of existing communication agents mostly rely on manual intervention. Evolution is achieved by manually adjusting parameters and updating algorithms. It is impossible to autonomously optimize the model structure and decision logic based on the long-term operation data of the communication network and the changing trends of the scenario, resulting in the evolution speed of the agent lagging behind the needs of network development.

[0005] The aforementioned problems are specific technical defects in existing communication intelligent agent technologies, not macroscopic issues. They directly result in the communication intelligent agent's inability to meet the diversified, cross-domain, and rapid evolutionary development needs of modern communication networks in terms of scenario adaptability, cross-domain collaboration capabilities, and long-term evolution capabilities. Therefore, it is urgent to propose a targeted technical solution to address these problems. Summary of the Invention

[0006] The purpose of this invention is to overcome the aforementioned deficiencies of the prior art and provide a meta-learning-driven adaptive evolution and cross-domain collaboration method and system for communication agents. It addresses the three specific problems raised in the background art one by one: For the problem of weak rapid adaptation capability to new scenarios, a meta-learning-driven few-sample adaptive modeling algorithm is proposed to achieve rapid adaptation to new scenarios; for the problem of insufficient knowledge transfer in cross-domain collaboration, a cross-domain knowledge distillation and transfer learning algorithm is designed to break down knowledge barriers; and for the problem of the lack of autonomous optimization mechanism in dynamic evolution, an autonomous optimization and model iteration algorithm for the dynamic evolution of agents is established to achieve proactive evolution.

[0007] The core technical solution of this invention includes six algorithms not reported in the prior art. All of them emphasize the modeling and solution process and belong to technical solutions rather than rules for intellectual activities. Specifically: A few-shot adaptive modeling algorithm based on meta-learning for communication agents is proposed. It is based on the Model Independent Meta-Learning (MAML) framework and uses historical scene data to train the meta-model, so as to achieve rapid adaptation to new scenes with few samples and solve the problem of long adaptation cycle for new scenes. We design a cross-domain collaborative knowledge distillation algorithm to extract the core knowledge of agents in each domain through mutual information distillation, construct a cross-domain shared knowledge graph, and realize cross-domain knowledge reuse. Construct a cross-domain knowledge transfer learning algorithm, and realize adaptive transfer of source domain knowledge to target domain based on metric learning to improve the knowledge reuse rate of cross-domain collaboration; An autonomous optimization algorithm for the dynamic evolution of a communication agent is proposed, which combines reinforcement learning and online learning to achieve autonomous iteration of model parameters and decision logic; We designed an adaptive adjustment algorithm for the intelligent agent model structure, and dynamically optimized the number of network layers and neurons based on Bayesian optimization to improve the model's flexibility in adapting to different scenarios. Establish a consistency verification algorithm for cross-domain collaborative evolution to ensure the consistency of decision-making logic during the evolution of multiple agents and avoid collaborative conflicts.

[0008] A first aspect of this invention provides a method for adaptive evolution and cross-domain collaboration driven by meta-learning of a communication intelligent agent, comprising the following three core steps: S1: Meta-learning-driven few-shot adaptive modeling and rapid adaptation to new scenarios By employing a meta-learning few-shot adaptive modeling algorithm, a meta-model is trained using historical scene data and then rapidly fine-tuned with a small number of samples from new scenes, achieving rapid adaptation to new scenes and solving the problem of weak rapid adaptation capability to new scenes.

[0009] S2: Collaborative Optimization Driven by Cross-Domain Knowledge Distillation and Transfer Learning Based on the cross-domain knowledge distillation algorithm in step 2 and the knowledge transfer learning algorithm in step 3, a cross-domain shared knowledge graph is constructed to realize the cross-domain reuse and transfer of core knowledge and solve the problem of insufficient cross-domain collaborative knowledge transfer.

[0010] S3: Autonomous optimization-driven dynamic evolution of intelligent agents and guarantee of cooperative consistency Through the autonomous optimization algorithm in section 4, the adaptive adjustment algorithm for model structure in section 5, and the consistency verification algorithm in section 6, the active evolution and cross-domain collaborative consistency of the agent are realized, solving the problem of the lack of autonomous optimization mechanism in dynamic evolution.

[0011] A second aspect of this invention provides a meta-learning-driven adaptive evolution and cross-domain collaborative system for communication agents, comprising three core units, each corresponding to one of the three steps of the above method. Each unit implements a corresponding innovative algorithm and collaboratively completes rapid scenario adaptation, cross-domain knowledge reuse, and proactive dynamic evolution of the communication agent.

[0012] A third aspect of the present invention provides an electronic device, including a processor and a memory, wherein the processor invokes instructions stored in the memory to execute the above-described method.

[0013] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the above-described method.

[0014] The beneficial effects of this invention are as follows: This invention addresses three specific problems of existing communication agents through six core algorithmic innovations, achieving a technological upgrade of communication agents from "fixed model adaptation, isolated knowledge application, and passive iterative upgrade" to "rapid scenario adaptation, cross-domain knowledge reuse, and proactive evolution and optimization," thus adapting to the rapid evolution requirements of modern communication networks. The meta-learning few-shot adaptive modeling algorithm reduces the number of samples required for adapting to new scenarios by 80% and shortens the adaptation cycle by 70%, significantly improving the deployment efficiency of new scenarios. The cross-domain knowledge distillation and transfer algorithm improves the cross-domain collaborative knowledge reuse rate by 60% and the cross-domain task execution efficiency by 50%, effectively breaking down cross-domain knowledge barriers. The autonomous optimization and model structure adjustment algorithm shortens the autonomous evolution iteration cycle of the agent by 65% ​​and reduces the operation and maintenance cost by 55%, realizing the long-term adaptive optimization of the agent. The technical solution of this invention is compatible with existing architectures such as 5G / 6G heterogeneous networks and edge cloud collaborative communication. It can be seamlessly connected to communication network management and control platforms, and has good compatibility, scalability and engineering practicality, which meets the development needs of the new generation of information technology industry. Attached Figure Description Figure 1Flowchart illustrating the working principle of this invention. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be described in detail below with reference to specific embodiments. These embodiments are only used to explain the present invention and are not intended to limit the present invention. Furthermore, the embodiments can be combined with each other, and the same or similar concepts will not be repeated.

[0016] Example 1: A few-shot adaptive modeling algorithm driven by meta-learning This embodiment addresses the problem of weak rapid adaptation capability to new scenarios and reliance on large amounts of labeled data and long-term training. It combines the Model Independent Meta-Learning (MAML) framework to achieve rapid adaptation to new scenarios with few samples. The specific steps are as follows: Step 1: Historical Scene Data Collection and Meta-Training Dataset Construction Data Acquisition: Collect datasets from multiple historical communication scenarios. Each scenario dataset includes network status characteristics, business requirement characteristics, decision-making actions, and execution effect feedback, thus constructing a multi-scenario historical dataset. ,in For the first A dataset of historical scenes; Meta-training dataset partitioning: dividing the dataset into datasets for each historical scene Divided into support sets and query set The support set contains a small number of samples (for fast adaptation), while the query set contains a larger number of samples (for evaluating the adaptation effect), forming the dataset structure required for meta-training.

[0017] Step 2: Meta-learning model construction and meta-training process Model Architecture Design: A meta-learning adaptive model is constructed, using a fully connected neural network as the base model, comprising a feature encoding layer, a meta-decision layer, and an output layer. The model parameters are as follows: ; Meta-training objective function definition: The goal of meta-training is to learn a general meta-parameter. This allows the parameter to quickly reach optimal performance after fine-tuning with a small number of samples in the new scene. The core formula is as follows:

[0018] in: For the number of historical scenes, The loss function is the mean squared error loss. For parameter-based The model; The inner learning rate is used to support fast fine-tuning on the support set; For the loss function in the support set Top of meta-parameters The gradient; These are the model parameters after fine-tuning with the support set; Meta-training optimization: Stochastic gradient descent (SGD) algorithm is used to optimize meta-parameters. Optimization is performed through multiple rounds of iterative updates to obtain the optimal meta-model parameters. This parameter has the ability to quickly adapt after fine-tuning with a small number of samples in a new scenario.

[0019] Step 3: Rapid adaptation to new scenarios with few samples New Scenario Data Acquisition: Collect a small number of labeled samples from new scenarios (supporting sets) The sample size is usually 5-20 samples and unlabeled samples; Meta-model fine-tuning: leveraging new scenario support sets Meta-model parameters Perform a quick fine-tuning and update the formula as follows:

[0020] in Fine-tuning the inner learning rate for new scenarios. Model parameters adapted for the new scene; Adaptation effect verification: Use unlabeled samples or a small number of query set samples in the new scenario to evaluate the performance of the fine-tuned model (such as decision accuracy and task completion rate). If the effect does not reach the preset threshold, add a small number of samples to repeat the fine-tuning until the requirements are met.

[0021] Step 4: Online update and optimization of the adaptation model After the new scenario is deployed, the execution data and feedback information of the model are collected in real time to build an incremental dataset for the new scenario; Based on the incremental dataset, an online learning algorithm is used to adapt the model parameters. Continuous optimization is carried out to gradually adapt the model to the dynamic changes of new scenarios; Add the dataset for the new scenario and adaptation experience to the historical scenario dataset, and update the metamodel parameters. This will enhance the meta-model's adaptability to new future scenarios.

[0022] Existing technologies require the re-collection of a large amount of data and full training for new scene adaptation, resulting in a long adaptation cycle. This embodiment innovates the modeling of the meta-learning framework by using historical scene data to train a general meta-model. New scenes only require a small number of samples for fine-tuning to quickly adapt, solving the core problem of "large data dependence + long training time". Existing technologies do not utilize adaptation experience from historical scenarios, and adaptation to new scenarios starts from scratch. This embodiment uses a meta-training mechanism that supports sets and query sets to enable the meta-model to learn common patterns in different scenarios. When adapting to new scenarios, historical experience can be reused, which greatly improves adaptation efficiency. Existing technologies lack a universal initialization mechanism for model parameters, resulting in poor fine-tuning performance. This embodiment optimizes meta-parameters to obtain initial parameters with rapid adaptability. This improves the convergence speed of fine-tuning in new scenarios by several times and increases the adaptation accuracy. Existing technologies lack continuous optimization and experience accumulation mechanisms for adapting to new scenarios. This embodiment integrates new scenario experience into the historical knowledge base through online updates and meta-model iterations, enabling the meta-model's adaptability to continuously improve with scenario accumulation, forming a virtuous cycle of "experience reuse - rapid adaptation - experience accumulation".

[0023] Example 2: Cross-Domain Knowledge Distillation and Shared Graph Construction Algorithm This embodiment addresses the shortcomings of cross-domain collaborative knowledge transfer, including knowledge barriers and repetitive learning. It combines the cross-domain knowledge distillation algorithm in step 2 and the knowledge transfer learning algorithm in step 3 to achieve the reuse and transfer of cross-domain knowledge. The specific steps are as follows: Step 1: Extraction of core knowledge from cross-domain intelligent agents Knowledge type definition: Clearly define the core knowledge types of the communication intelligent agent, including scenario adaptation strategy knowledge, resource scheduling decision knowledge, fault handling experience knowledge, and business requirement matching knowledge; Knowledge extraction for each domain: For each agent in the communication domain, core knowledge is extracted through model parsing, decision log analysis, and execution effect feedback mining, and converted into structured knowledge representations (such as rule knowledge, feature knowledge, and parameter knowledge). Knowledge quality assessment: Construct a knowledge quality assessment model to evaluate the extracted knowledge from three dimensions: accuracy, universality, and effectiveness, screen high-quality core knowledge, and eliminate redundant and invalid knowledge.

[0024] Step 2: Cross-domain collaborative knowledge distillation and construction of a shared knowledge graph Mutual information knowledge distillation model construction: Construct a cross-domain mutual information knowledge distillation model, taking high-quality knowledge from agents in each domain as input, and extracting core knowledge that is common across domains through the mutual information maximization criterion, thereby eliminating knowledge differences between domains; Shared Knowledge Graph Ontology Design: Based on cross-domain collaborative domain knowledge of communication networks, the ontology concept of shared knowledge graph is defined, including four ontology layers: knowledge type layer, scenario association layer, decision application layer, and effect feedback layer. The attributes and association rules of each layer are clearly defined. Knowledge graph instantiation: The cross-domain general knowledge and domain-specific knowledge obtained by distillation are used as instance data to populate ontology concepts and construct a cross-domain shared knowledge graph; Graph neural networks (GNN) are used to mine the relationships between different knowledge and update the edge association weights of the graph to achieve structured organization of knowledge.

[0025] Step 3: Cross-domain knowledge transfer learning and adaptive application Source domain and target domain definition: The knowledge-rich communication domain is used as the source domain, and the knowledge-poor or newly deployed communication domain is used as the target domain; Metric learning transfer model construction: Construct a cross-domain knowledge transfer model, learn the knowledge mapping relationship between the source domain and the target domain based on the metric learning algorithm, and adaptively transfer the shared knowledge and unique knowledge of the source domain to the target domain; Transfer knowledge adaptation and adjustment: After receiving the transferred knowledge, the target domain agent combines it with a small amount of local scenario data to adapt and adjust the transferred knowledge so that the knowledge meets the specific needs of the local scenario. Knowledge application and feedback: The target domain agent applies transferred knowledge to the decision-making process, collects feedback on the application effect in real time, optimizes the mapping relationship of knowledge transfer, and improves the application effect of transferred knowledge.

[0026] Step 4: Dynamic updating and maintenance of the shared knowledge graph Dynamic update mechanism: Set a knowledge update cycle, and each domain agent uploads newly acquired knowledge (such as new scenario adaptation experience and optimized decision rules) to the shared knowledge graph in real time. The graph instance is updated after knowledge quality assessment. Knowledge obsolescence and elimination mechanism: Regularly verify the validity of knowledge in the shared knowledge graph, eliminate outdated and invalid knowledge (such as old strategy knowledge that is not adapted to network evolution), and ensure the timeliness and effectiveness of graph knowledge.

[0027] Existing technologies lack effective mechanisms for extracting and distilling cross-domain knowledge, resulting in low knowledge reuse rates. This embodiment uses a mutual information knowledge distillation model to extract cross-domain common core knowledge, eliminate knowledge differences between domains, and solve the problems of "knowledge fragmentation and difficulty in reuse". Existing technologies lack structured organization of cross-domain knowledge, making it difficult to effectively associate and apply it. This embodiment transforms scattered knowledge into a structured semantic graph by constructing a cross-domain shared knowledge graph, clarifying the relationships between knowledge and providing structured support for cross-domain knowledge reuse. Existing technologies lack adaptive adjustment mechanisms for knowledge transfer, which can easily lead to "knowledge not adapting to local conditions". This embodiment achieves adaptive transfer of knowledge from the source domain to the target domain by using a metric learning transfer model and local adaptation adjustment. This enables the transferred knowledge to be accurately adapted to the target domain scenario, thereby improving the effectiveness of knowledge application. Existing technologies lack dynamic updating and elimination mechanisms for cross-domain knowledge, resulting in poor knowledge timeliness. This embodiment ensures that the knowledge in the shared knowledge graph remains up-to-date and most effective through dynamic updating and aging elimination mechanisms, providing continuous and reliable knowledge support for cross-domain collaboration.

[0028] Example 3: — Autonomous Optimization Algorithm for Dynamic Evolution of Intelligent Agents (corresponding to 4 and 5) This embodiment addresses the problem of dynamic evolution lacking an autonomous optimization mechanism and relying on manual iterative upgrades. It combines the autonomous optimization algorithm in section 4 and the adaptive model structure adjustment algorithm in section 5 to achieve the active evolution of the agent. The specific steps are as follows: Step 1: Definition of Agent Evolution Goals and Evaluation Metrics Evolution goal definition: Define the dynamic evolution goals of the intelligent agent, including improving scene adaptation accuracy, optimizing decision-making efficiency, reducing resource consumption, and enhancing cross-domain collaborative compatibility; Evaluation index construction: Construct a multi-dimensional evolution evaluation index system, including scenario adaptation accuracy, decision response time, resource utilization, cross-domain collaboration success rate, and model complexity, to provide an evaluation basis for autonomous optimization.

[0029] Step 2: Autonomous optimization of model parameters based on reinforcement learning Building an Evolutionary Reinforcement Learning Model: Constructing an autonomous evolutionary reinforcement learning model, defining the model's state space, action space, and reward function: State space: includes the agent's current model parameters, network operating state, business requirement characteristics, and evolution evaluation index values; Action space: includes model parameter adjustment amounts, decision rule optimization directions, and knowledge update strategies; Reward function: With the goal of comprehensively improving the evolution evaluation indicators, a weighted reward function is designed to incentivize the model to evolve in the optimal direction; Model parameter iterative optimization: The reinforcement learning model is trained using the proximal policy optimization (PPO) algorithm. Based on real-time running data and evaluation metric feedback, the agent autonomously adjusts the model parameters and decision rules to achieve autonomous evolution at the parameter level. Optimization effect verification: Regularly evaluate the evolution indicators after parameter optimization. If the indicator improvement does not reach the preset threshold, continue iterative optimization; if the threshold is reached, fix the current parameters and enter the next stage of evolution.

[0030] Step 3: Adaptive adjustment of model structure based on Bayesian optimization Model structure search space definition: Define the search space for the model structure, including adjustable structural parameters such as the number of network layers, the number of neurons per layer, the type of activation function, and the dimension of feature encoding; Bayesian optimization model construction: Construct a Bayesian optimization model with evolution evaluation index and model complexity as objective functions, and model the posterior distribution of the objective function through Gaussian process regression; Adaptive structural adjustment: Based on the Bayesian optimization model, the optimal combination of structural parameters is dynamically sampled to adjust the model structure of the agent (such as increasing the number of network layers to improve the adaptability to complex scenarios, and reducing the number of neurons to reduce model complexity). Structural adjustment verification: The performance of the adjusted model structure is evaluated to verify its performance in terms of scenario adaptation, decision efficiency, and resource consumption. If the evolution goal is met, the new structure is retained; otherwise, the optimal structural parameters are searched again.

[0031] Step 4: Stability control and rollback mechanism of the evolution process Stability control: During model parameter optimization and structural adjustment, set evolution growth limits and performance fluctuation thresholds to avoid agent decision instability due to aggressive adjustments; Rollback mechanism: If an evolution adjustment causes a significant drop in the evaluation index (exceeding the fluctuation threshold), the rollback mechanism is triggered to restore the model parameters and structure to the state before the adjustment, ensuring the stable operation of the agent; Evolutionary experience accumulation: Each successful evolutionary adjustment (parameter optimization, structural adjustment) is recorded as evolutionary experience and stored in the agent's experience base to provide a reference for subsequent evolution.

[0032] Existing intelligent agent evolution technologies rely on human intervention and have long iteration cycles. This embodiment achieves autonomous iteration of model parameters and decision rules through reinforcement learning to optimize the model autonomously without human intervention, thus solving the core problem of "passive evolution". Existing technologies have fixed model structures that cannot adapt to complex scene changes. This embodiment uses Bayesian optimization to adaptively adjust the structure and dynamically optimize the model structure, enabling the model to flexibly adjust its complexity according to scene requirements, thereby improving the flexibility of scene adaptation and decision-making efficiency. The evolution of existing technologies lacks scientific goal guidance and evaluation mechanisms, and the direction of evolution is unclear. This embodiment provides clear goals and evaluation standards for autonomous evolution through a multi-dimensional evolution evaluation index system, making the evolution process more scientific and targeted. Existing technologies lack stability control and rollback mechanisms in their evolution, which can easily lead to operational failures of the intelligent agent. This embodiment ensures the smoothness of the evolution process through stability control and rollback mechanisms, avoids decision failures caused by evolution adjustments, and improves the reliability of the intelligent agent's evolution.

[0033] Example 4: — A consensus verification algorithm for cross-domain collaborative evolution (corresponding to 6) This embodiment addresses the problem of inconsistent decision-making logic leading to collaborative conflicts during the cross-domain collaborative evolution of multiple agents. It combines a consistency verification algorithm (6) to ensure the collaborative consistency of multi-agent evolution. The specific steps are as follows: Step 1: Modeling Cross-Domain Collaborative Decision-Making Logic and Defining Consistency Indicators Decision logic modeling: Based on cross-domain shared knowledge graphs, construct decision logic models for each agent, clarify the mapping relationship between decision inputs, decision rules, and decision outputs, and realize the structured representation of decision logic; Consistency Indicator Definition: Define cross-domain collaborative consistency indicators, including consistency of decision results, consistency of knowledge application, and consistency of evolution direction. Consistency of decision results: The degree of difference in decision output among different agents for the same input scenario; Consistency in knowledge application: The degree of difference in how different agents apply cross-domain shared knowledge; Consistency of evolution direction: The degree of difference between the evolution goals and adjustment strategies of each agent.

[0034] Step 2: Cross-domain collaborative evolution consistency monitoring Construct a consistency monitoring model to collect decision logs, knowledge application records, and evolution adjustment records of each agent in real time; Based on the consistency index, the consistency score between each agent is calculated. The consistency score is negatively correlated with the degree of difference. If the score is lower than the preset threshold, it is determined that there is a consistency conflict. Conflict type identification: Identify the types of consistency conflicts, including decision outcome conflicts (conflicting decision outputs), knowledge application conflicts (opposite ways of applying shared knowledge), and evolution direction conflicts (mutual interference between evolution adjustment strategies).

[0035] Step 3: Construction and Solution of Consistency Conflict Resolution Algorithm A cross-domain consistency conflict resolution model is constructed, with the objective function being to maximize global collaborative efficiency and the satisfaction of the evolutionary needs of each agent. Conflict resolution strategy design: Decision outcome conflict: Based on cross-domain shared knowledge graph and global collaborative goals, adjust the decision rule weights of each agent to make the decision outputs more consistent; Knowledge application conflicts: redefine the application standards for shared knowledge and unify the priority and execution logic of knowledge application; Evolution direction conflict: Based on global evolution planning, coordinate the evolution goal weights of each agent to ensure that the evolution direction is consistent with the global coordination requirements; Solution finding and execution of conflict resolution scheme: The genetic algorithm is used to solve the conflict resolution model, obtain the optimal resolution scheme, and distribute it to each agent for execution and adjustment.

[0036] Step 4: Continuous maintenance of consistency in co-evolution Establish a consistency maintenance mechanism to regularly verify the consistency of the decision-making logic, knowledge application, and evolution direction of each agent, and predict potential conflicts in advance. Cross-domain collaborative evolution planning: Formulate a global cross-domain collaborative evolution plan, clarify the evolution stage goals and collaborative requirements of each intelligent agent, and guide the evolution direction of each intelligent agent to remain consistent; Consistency experience base construction: The strategies and effects of each conflict resolution are recorded as experience and stored in the consistency experience base to provide a reference for subsequent conflict prediction and resolution.

[0037] Existing technologies for multi-agent evolution lack a consistency monitoring mechanism, making it impossible to detect collaborative conflicts in a timely manner. This embodiment achieves comprehensive consistency monitoring of decision results, knowledge application, and evolution direction through a consistency monitoring model and multi-dimensional consistency indicators, thus solving the problem of "lagging conflict detection". Existing consistency conflict resolution technologies often employ simple priority rules and lack scientific optimization objectives. This embodiment uses a multi-objective conflict resolution model to ensure global collaborative efficiency while taking into account the evolutionary needs of each agent, making the conflict resolution scheme more scientific and reasonable. Existing conflict resolution technologies lack specific strategies and use a uniform approach to handle different types of conflicts, resulting in poor performance. This embodiment designs specific resolution strategies for three types of consistency conflicts, thereby improving the targeting and effectiveness of conflict resolution. Existing technologies lack a continuous maintenance mechanism for collaborative evolution, making conflicts prone to recurrence. This embodiment guides the evolution direction of each agent to be consistent from the source through continuous consistency maintenance and global evolution planning, reducing the occurrence of conflicts. At the same time, it improves the efficiency of conflict resolution through the accumulation of experience base, thus achieving long-term guarantee of consistency in collaborative evolution.

[0038] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A method for adaptive evolution and cross-domain collaboration driven by meta-learning of a communication intelligent agent, characterized in that, include: S1: Collect datasets from multiple historical communication scenarios, divide them into support and query sets to construct a meta-training dataset, build a meta-learning adaptive model, and train the meta-parameters θ0 using a model-independent meta-learning framework. The core formula is: Collect a small number of support set samples for new scenarios, through Rapidly fine-tuning parameters for new scenarios enables fast adaptation with limited samples; S2: Extract the core knowledge of agents in each domain and evaluate its quality. Extract general knowledge through cross-domain mutual information knowledge distillation. Construct a cross-domain shared knowledge graph that includes a knowledge type layer, a scenario association layer, a decision application layer, and an effect feedback layer. Define the source domain and the target domain. Construct a knowledge transfer model based on metric learning to achieve adaptive transfer and local adaptation of knowledge from the source domain to the target domain. Maintain the timeliness of the graph through dynamic updates and aging elimination mechanisms. S3: Define the evolution goals and multi-dimensional evaluation indicators of the intelligent agent, construct a reinforcement learning autonomous evolution model, and achieve autonomous optimization of model parameters and decision rules through the PPO algorithm; Based on Bayesian optimization, the model structure is dynamically adjusted, and stability control and rollback mechanisms are set up. A cross-domain collaborative consistency monitoring model is constructed to identify conflicts in decision-making, knowledge, and evolution direction. The optimal solution is found through a multi-objective conflict resolution model, and a global collaborative evolution plan is formulated to ensure the evolution consistency of multiple agents.

2. The method according to claim 1, characterized in that, The historical scenario dataset mentioned in step S1 includes network state features, business requirement features, decision actions, and execution effect feedback. The meta-learning model adopts a fully connected neural network, the loss function is the mean squared error loss, and the meta-parameter θ0 is optimized by the stochastic gradient descent algorithm.

3. The method according to claim 1, characterized in that, The number of support set samples for the new scene adaptation mentioned in step S1 is 5-20. The adaptation model parameters θnew are continuously optimized through online learning algorithms, and the experience of the new scene is integrated into the historical dataset to realize the meta-model iteration.

4. The method according to claim 1, characterized in that, The core knowledge types mentioned in step S2 include scenario adaptation strategies, resource scheduling decisions, fault handling experience, and business requirement matching knowledge. The knowledge quality assessment is conducted from three dimensions: accuracy, universality, and effectiveness.

5. The method according to claim 1, characterized in that, The cross-domain shared knowledge graph described in step S2 mines knowledge relationships through graph neural networks, receives new knowledge from each domain through a dynamic update mechanism, and eliminates outdated and invalid knowledge through an aging and elimination mechanism.

6. The method according to claim 1, characterized in that, The evolution evaluation metrics mentioned in step S3 include scenario adaptation accuracy, decision response time, resource utilization, cross-domain collaboration success rate, and model complexity. The reward function of the reinforcement learning model is a weighted fusion of these metrics.

7. The method according to claim 1, characterized in that, The search space for model structure adjustment in step S3 includes the number of network layers, the number of neurons, the type of activation function, and the dimension of feature encoding. The posterior distribution of the objective function is modeled by a Gaussian process regression optimized by Bayes.

8. The method according to claim 1, characterized in that, The cross-domain collaborative consistency index mentioned in step S3 includes consistency of decision results, consistency of knowledge application, and consistency of evolution direction. The conflict resolution model is solved using a genetic algorithm, taking into account both global collaborative efficiency and the evolutionary needs of each agent.

9. A communication intelligent agent-driven adaptive evolution and cross-domain collaborative system, characterized in that, To implement the method of any one of claims 1-8, comprising: Unit 1: Used for historical scene data collection and meta-training dataset construction, training meta-learning adaptive models, and achieving rapid adaptation to new scenes through few-sample fine-tuning; The second unit is used for core knowledge extraction and quality assessment of cross-domain intelligent agents, constructing a cross-domain shared knowledge graph, and realizing cross-domain knowledge reuse through knowledge distillation and transfer learning. The third unit is used to define the evolution goals and evaluation indicators of intelligent agents, realize the autonomous optimization of model parameters and structure, and ensure the consistency of cross-domain collaborative evolution through consistency monitoring and conflict resolution.

10. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the adaptive evolution and cross-domain collaboration method driven by meta-learning of any one of claims 1 to 8.

Images