结构自适应自演化的高级机器学习方法研究与进展  

Research and Progress of Advanced Machine Learning Methods for Structure Adaptive Self-Evolution

作　　者：李玺 高佳 励雪巍[1] 王祥丰 金博 陈喜群[4] 钱徽[1] LI Xi;GAO Jia;LI Xuewei;WANG Xiangfeng;JIN Bo;CHEN Xiqun;QIAN Hui(College of Computer Science and Technology,Zhejiang University,Hangzhou 310027;School of Automation and Software Engineering,Shanxi University,Taiyuan 237016;Software Engineering Institute,East China Normal University,Shanghai 200062;College of Civil Engineering and Architecture,Zhejiang University,Hangzhou 310027)

机构地区：[1]浙江大学计算机科学与技术学院,杭州310027 [2]山西大学自动化与软件学院,太原237016 [3]华东师范大学软件工程学院,上海200062 [4]浙江大学建筑工程学院,杭州310027

出　　处：《中国基础科学》2022年第3期47-53,共7页China Basic Science

基　　金：科技创新2030项目(2020AAA0107400);国家自然科学基金项目(U20A20222);浙江省杰出青年科学基金项目(LR19F020004)

摘　　要：针对当前机器学习面临的结构机制认知不足、学习方式自主性不强、模型推理适应性不高和学习形态封闭单一等重大挑战,本文针对如下3个关键科技问题进行阐述:(1)多任务知识迁移和模型动态自适应学习机制;(2)协同多智能体系统的分布式自组织和自演化学习机制;(3)面向结构化数据在线学习的动态建模与评价。建立面向多任务知识迁移的自适应元学习框架;提出基于层次共识的多智能体自组织协同学习与自演化机制,构建基于弹性关联拓扑的智能体分布式学习方法;搭建基于结构化、自组织和自演化学习框架的仿真验证系统。为人工智能基础理论研究提供新的理论方法和计算模式,支持媒体语义学习、实时计算和知识发现等重大应用。In view of the current machine learning facing major challenges,such as insufficient cognition of structure mechanism,weak autonomy of learning mode,low adaptability of model reasoning,and unitary closedform learning mode,this paper addresses the following three key scientific and technological problems:1)multitask knowledge transfer and dynamically adaptive model learning mechanism;2)distributed self-organizing and self-evolving learning mechanism of cooperative multiagent systems;3)dynamic modeling and evaluation for online learning of structured data.An adaptive meta-learning framework for multi-task knowledge transfer is established;a multi-agent self-organizing cooperative learning and self-evolving mechanism based on hierarchical consensus is proposed,and an agent distributed learning method based on elastic association topology is constructed;a simulation verification system based on structured,self-organizing and self-evolving learning framework is built.It provides a new theoretical methodology and computing mode for the basic research on artificial intelligence,and gives rise to several important applications such as media semantic learning,real-time computing and knowledge discovery.

关 键 词：学习机制 自适应 自组织 自演化 动态建模 仿真验证 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]