Research On Reinforcement Learning In Agent Model Of Decision Making Simulation System

Reinforcement learning is an unsupervised machine learning technology, whichcan find the optimal solution or near optimal solution through trial and error method,and can implement online learning in the dynamic environment. Therefore,Reinforcement learning is considered as an ideal technology to constitute intelligentAgent. Complex environment and state of uncertainty are characteristics ofDecision-making Simulation System, and how to apply the Reinforcement Learning tothe decision-making simulation field is hot spot and difficult point in current research.This paper mainly studies Reinforcement Learning method and applied tomulti-Agent Decision-making Simulation System. First, this paper gives an Agent-baseddecision-making simulation system framework according to the thought of hierarchicallearning and decision-making. Then a cognitive Agent model based on ReinforcementLearning mathod is present, in which the prediction mechanism of the environmentchanges is introduced. Through the introduction, Agent has forward-looking predictioncapabilities of predicting the environmental changes to some extent. Based on theanalysis of available reinforcement learning algorithms, an improved single-Agent Qlearning algorithm is given, so that each Agent in the system can learn by itself asneeded. Subsequently, this paper analyzes the credit assignment problem in multi-Agentreinforcement learning and proposes a multi-Agent union Q-Learning algorithm toimprove the learning efficiency of multi-Agent system.Finally, validations of the cognitive Agent model and the Reinforcement Learningalgorithm are given through the experiment of certain a decision-making simulationsystem. The experimental results show that the proposed cognitive Agent model can beapplied to decision-making simulation systems effectively and the reinforcementlearning algorithms improved the Agents’ learning efficiency.