Research On Multi-agent Collaborative Perception Semi-supervised Online Evolutive Learning

With the development of artificial intelligence and edge intelligence technology,various intelligent perception services provide a lot of convenience for our lives.However,at this stage,most perception and recognition services still rely on a large number of manual annotations for model training and tuning.As the perception environment continues changing,the generalization of various artificial intelligence models will be significantly affected.Online evolutive learning(OEL)will be an effective solution to reduce manual annotation and enable various intelligent services to adapt to changes in the perception environment autonomously.OEL performs collaborative discrimination and knowledge sharing on perceptual data through multiagents,which can make edge models continuously adapt to data distribution changes.Semi-supervised learning(SSL)can use a small amount of labeled data and mine effective generalization information from unlabeled data to constantly improve the performance of specific recognition tasks without labeling,which will be an effective way to achieve OEL.The main goal of this research is to enable the model to autonomously adapt to changes in the sensing environment through information interaction,continuously improve the generalization when performing specific tasks,and realize multi-agent collaborative perception Online Evolutive Learning.This paper will take SSL as the fundamental method and first solve several critical problems affecting SSL’s learning efficiency and stability.Afterward,the collaborative perception method of multi-agent models in a multi-view perception environment is studied.Then for more practical perception scenarios,the online evolutionary learning problem of each agent model facing non-i.i.d.perceptual data is studied.The specific research contents are as follows:To improve the stability and training efficiency of SSL and thus strengthen its application ability in OEL scenarios,I propose the adaptive weighted losses and distribution approximation(AWL&DA)semi-supervised learning method.The consistency-based semi-supervised learning method is prone to the class imbalance problem in the training process due to the lack of constraints on class-wise learning progress differences.By evaluating the class-wise learning progress of each data during the training process,different training contributions are adaptively allocated to each data category,the category imbalance problem in the SSL training process is reduced,and the training stability is improved.By approximating the bidirectional distribution between the low-confidence prediction distributions during the training process,the proposed method significantly improves the convergence efficiency of the SSL model.Experimental analysis results show that our approach surpasses the performance of many top SSL algorithms in several representative experiments.At the same time,when facing an extreme labeled data scarcity situation,such as in the CIFAR-10-40-labels experiment,our approach significantly improves the convergence speed and achieves state-of-the-art performance.To solve the OEL problem in the multi-view agent perception environment and to improve the shortcomings of the existing multi-view or multi-model SSL algorithms in terms of application scope and construction of differentiated discriminant models,I propose Multi-view Agent Collaborative Perception(MACP)semi-supervised online evolutive learning method.Through the self-supervised initialization of different view models,each agent model can obtain different feature extraction patterns in the initial training stage.Through the discriminative information fusion process of the multi-view model,more reliable annotation information is mined from the prediction of multi-view perceptual information.By introducing additional parameter constraints between different view models in the training process,the models can maintain relatively independent discriminative capabilities throughout the training process,improving the stability of the overall learning system.The proposed algorithm surpasses the performance of multiple multi-model and single-model SSL algorithms in experiments on various databases.In an ideal multi-view perception environment,MACP achieves convergence efficiency and performance that surpasses fully supervised learning methods.To solve the frequent changing data distribution problem in the practical perception environment and make up for the shortcomings of the general SSL algorithm in model interaction design,I propose the mutual match(MM)semi-supervised online evolutive learning method.Through soft-supervision consistency regularization,each agent model can learn more category-related information.Through the model predicted distribution sharing method of mutual interactive learning,each model can efficiently share the critical knowledge related to the discriminant target.As a result,the entire learning system can effectively adapt to distribution changes of perception data.Through the unsupervised sample mining process,each model can independently use reliable perception data and prediction results to expand the labeled data set continuously.In a unified perception environment with a large amount of non-i.i.d.data,the training efficiency and accuracy of the proposed algorithm surpass that of many representative SSL algorithms.The experimental analysis results show that the unified framework of the MM method is more suitable for solving the data distribution problem in the OEL environment,so as to realize the real-time interaction and evolutionary learning of each agent model and continuously improve the adaptability of the OEL system to environmental changes.With the evolution of efficient wireless communication technology and lightweight high-performance computing chips,online evolutive learning research in the field of edge intelligence will have broad development prospects.With the blessing of OEL,more artificial intelligence applications that solve practical perception tasks and provide real-time services can reduce or eliminate the dependence on cloud models and manual annotations.Multi-agents can continuously self-adaptation to the current perception environment and realize more intelligent customized services.