MEBRL: Memory-Evolution-Based Reinforcement Learning Algorithm of MAS

Abstract

A memory-evolution-based MAS reinforcement learning algorithm (MEBRL) inspired by a psychology memory model is presented. 3 types of different memory stores are used in the design of the algorithm and Learning Automata is used in the processes of agent memory evolution. Through the memory evolution procedure, the agent in the MAS could make a proper decision and share its information indirectly. A multi-agent multi-resource stochastic system model is used to illustrate the performance of the algorithm, and the comparison of the memory-evolution-based MAS reinforcement learning algorithm and other MAS learning algorithm is given.