基于稀疏化核学习的智能网络分布式估计与动态优化研究

In the big data environment, there are three intertwined basic scientific problems need to intensive investigate in intelligence network, which include the mechanism of designing distributed estimation algorithm, the theoretical basis of constructing data sparsification algorithm, and the approach of realizing distributed optimization. By theoretical innovation and laboratory proofing, this project carries out research on distributed estimation and dynamic optimization with sparse kernel learning in intelligence network. The aims of the project are to solve the key problems of signal agent reinforcement learning based on sparse kernel method, distributed kernel adaptive filtering based on diffusion strategy, robust optimization based on Markov game, and the experiment platform development of the project. The main research contents are as follows: 1) Building online sparsification algorithm by kernel method, signal agent reinforcement learning is realized; 2) Based on diffusion strategy, the distributed kernel adaptive filtering is designed to online estimation about uncertain parameters of network; 3) According to the theory of Markov game, the robust optimization algorithm is constructed to explore a new way to solve Nash equilibrium and to investigate the robustness and optimality of intelligence network; 4) By designing computer simulation and experiment platform of this project, all research results not only are tested, but are provided technical support for development and designing of intelligence network in practical application.

大数据环境下，通过网络拓扑设计智能网络分布式估计算法的机制、构建数据稀疏化算法的机理、实现分布式优化的途径等基本科学问题需要进行深入研究。本项目通过理论创新和实验室验证，开展基于稀疏化核学习的智能网络分布式估计与动态优化研究。拟解决基于稀疏化核方法的智能体强化学习、基于扩散策略的分布式核自适应滤波、基于Markov对策的分布式鲁棒优化和智能网络分布式估计与动态优化实验平台开发等关键问题。主要研究内容包括：1) 引入核方法，在线构建稀疏化算法，实现单智能体强化学习；2) 基于扩散策略，设计分布式核自适应滤波算法、对网络不确定性进行在线估计；3) 基于Markov对策设计分布式鲁棒优化算法，对网络系统的鲁棒性和最优性进行研究，为获得Nash平衡解探索一种新的途径；4) 构建分布式估计与动态优化计算机仿真与实验平台，检验所有研究成果，为具有实际应用的智能网络系统的开发与设计提供技术支撑。

本项目从工程实际应用出发，基于在线稀疏化核学习、强化学习、多智能体协调控制、博弈理论、最优化理论、自适应滤波理论、现代数学等相关知识研究智能网络分布式估计与动态优化问题。基于Mercer核函数及其对应特征空间的非线性映射，在线构建核稀疏化算法，实现单智能体自适应学习和自适应滤波。并且进一步研究生成对抗神经网络数学模型及其在计算机视觉算法的主要实现形式。基于代数图论深入研究智能网络分布式优化实现机理，设计分布式学习和自适应滤波算法、对网络不确定性进行在线估计。基于博弈理论研究多智能体网络系统无模型强化学习，设计分布式鲁棒优化算法，对网络系统的鲁棒性和最优性进行研究，为获得Nash均衡探索一种新途径。基于水声传感器网路，深入研究水下传感器网络的功率控制、水声信号的采集传输和定位算法。设计智能网络分布式估计与动态优化计算机仿真与实验平台，通过理论分析、计算机仿真和实验室验证，检验所有研究成果，为具有实际应用的智能网络系统的开发与设计提供理论依据。