基于神经网络的Agent增强学习模型  

An Agent Reinforcement Learning Model Based on Neural Networks

作　　者：唐亮贵[1] 刘波[1] 唐灿[1] 程代杰[2] 

机构地区：[1]重庆工商大学计算机学院,重庆400067 [2]重庆大学计算机学院,重庆400030

出　　处：《计算机科学》2007年第11期156-158,297,共4页Computer Science

基　　金：重庆市重点科技攻关资助项目(CSTC;2005AC2090);重庆市自然科学基金(CSTC;2004BB2167;CSTC;2006BB2249);重庆市教委科技项目(KJ060704)

摘　　要：在深入分析Agent决策过程中状态与行为空间的迁移与构造的基础上,设计了Agent基于强化学习的最优行为选择策略和Agent强化学习的神经网络模型与算法,并对算法的收敛性进行了证明。通过对多Agent电子商务系统中Agent竞价行为的预测仿真实验,验证了基于神经网络的Agent强化学习算法具有良好的性能和行为逼近能力。This paper thoroughly analyzes the transfer and construction of the state-action space of the agent decision- making process, discusses the optimal strategy of agent＇s action selection based on Markov decision-making process, and designed a neural networks model for the agent reinforcement learning, and designed the agent reinforcement learning based on neural networks. By the simulation experiment of agent＇s bid price in Multi-Agent Electronic Commerce System, validated the Agent Reinforcement Learning Algorithm Based on Neural Networks has very good performance and the action impending ability.

关 键 词：AGENT 强化学习 神经网络 MARKOV决策过程 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]