Consistency And Excellence Of Interactive Multi-agent Learning Systems

Machine learning is becoming dominant in the field of artificial intelligence,with significant successes in computer vision,natural language processing,and multi-agent learning.However,as neural network structures become progressively more complex and application scenarios more widespread,fundamental problems such as the poor robustness of deep learning models,the inability to adapt to new tasks,and the difficulty in scaling up become increasingly apparent.In contrast,group behavior in nature tends to produce robust,adaptable,flexible,and easily scalable systems.Group intelligence is the emergence of group behavior from the interaction of many individuals,enabling species to perform tasks that each individual could not do alone.Therefore,the thesis will incorporate the thought of group intelligence into the learning system,focusing on the interaction rules,consistency,and excellence among agents.This thesis utilizes a systems science approach to research the network structure and information interaction rules among multi-agent from a systematic and holistic perspective,providing new ideas for exploring the mechanism of group intelligence emergence.It is shown that in the field of cognitive learning,each agent can be expressed as a different neural network.Through continuous interaction between agents,group behaviors that are superior to individual intelligence will emerge.There are various ways of interaction among agents,such as adversarial,cooperative,and alternating adversarial cooperation.In the thesis,interaction rules are used as the main line to carry out the theory and method of interactive learning in multi-agent learning systems in the field of cognitive learning.The main research contents and innovations of the thesis are summarized as follows:(1)A multi-agent distributed generative adversarial network is proposed and conditions for multi-agent adversarial learning to reach consensus are given.Generative adversarial networks are typical of adversarial learning.In the thesis,the mathematical nature of generative adversarial networks is revealed through conceptual dynamics.The generators and discriminators in generative adversarial networks are considered different classes of multi-agent.A multi-stage information interaction is established for the model,and the overall convergence goal is achieved based on the multi-agent system consistency theory and algorithms,thus showing that the goal of the multi-agent consistency algorithm is essentially the same as the Nash equilibrium of the generative adversarial network.Secondly,the existence of the solution of the Degroot model in the conceptual dynamical system proves that the generators and discriminators in the generative adversarial network can agree on the distribution function,thus obtaining a new sufficient necessary condition for the existence of Nash equilibrium in the generative adversarial network.Further,to solve the problem of multi-agent consistency in complex distributed systems,a new distributed generative adversarial network is proposed in the thesis.This network treats discriminators with relatively greater influence as leaders and generators as followers according to the degree of influence among agents.By analyzing whether there is a smooth distribution of Markov chains composed of multi-agent states,the conditions for a distributed generative adversarial network with multi-generator with leaders(the agents with a high relative degree of influence)and multi-discriminator(multi-agent)to reach agreement are given.The correctness of the theoretical results is verified on the MNIST handwritten digits,anime character avatars,and face image datasets.(2)A dynamic interaction learning model for multi-agent with dynamic interaction topology is proposed,and the model convergence conditions are given.Under the context of data privacy protection,this thesis focuses on how the multi-agent emerge as a group intelligence through dynamic cooperative interactions to obtain more accurate learning results.First,each site is considered an independent agent,with each site having its own private data and uneven data across sites.In this thesis,dynamic interaction rules are introduced to combine multiple sites for interaction learning,and the conditions for convergence of dynamic interaction rules are given.The dynamic interaction process does not require all the agents to participate,and the topology of the network composed of agents can be changed dynamically.Further,to address the problems of expensive communication,complex and variable communication environment,and data privacy and domain knowledge barriers in practical applications,a multi-agent dynamic interaction learning model is proposed in this thesis.The model considers how multi-agent can achieve better learning than the single agent in the presence of limited and unreliable information exchange under dynamically changing interaction topologies.The multi-agent dynamic interaction learning model achieves decentralization and can alleviate data privacy,knowledge barriers,and security issues in the communication process.The multi-agent dynamic interaction learning model achieves decentralization,which can alleviate data privacy and security issues during communication.Finally,this thesis applies the multi-agent dynamic interaction learning model to the CIFAR-10 and COVID-19 pneumonia image data classification tasks.The experimental results show that the multi-agent dynamic interaction learning results outperform the single-site learning.(3)The conditions for multi-agent learning to outperform its subset are proposed and mathematically proven theoretically.First,by minimizing the empirical risk,this thesis demonstrates that the quality of multi-agent learning is better than its subsets with a high probability as the number of agents increases.Theoretically,we analyze the reason why multi-agent learning outperforms individual learning and provide new ideas for researching the mechanism of group intelligence emergence.Further,this thesis constructs a multi-agent interaction learning model and applies it to time series prediction,language modeling,and semantic segmentation tasks.The experimental results show that with the increasing number of agents,the multi-agent learning results outperform their subsets with high probability.In addition,it is found that interaction can effectively improve multi-agent learning results.The multi-agent interaction learning results are better than multi-agent learning that only fuses the final results.