Agent Learning Game Strategy Of The Mas Environment

In recent years, Agent and Multi-Agent System (MAS) is growing up to be one of the most important techniques in the practical research on Artificial Intelligence and intelligent software in distributed computing environment. Multi-agent games are becoming an increasingly prevalent formalism for the study of electronic commerce and auctions. The speed at which transactions can take palace and the growing complexity of electronic marketplaces makes the study of computationally simple agents an appealing direction.Agents in a MAS typically operate in large, complex, open, dynamic and unpredictable environments. The optimal policy at any moment depends on the policies of the other agents and so creates a situation of learning a moving target. Multi-agent learning is not only an intersection of Distributed Artificial Intelligence (DAI) and Machine Learning (ML), but also an area where ML and Game Theory meet.In competition environment, a satisfactory multi-agent learning algorithm should, at a minimum, have rationality and convergence. On the other hand, the algorithm should make the agent beat other fair opponents. Recently, some algorithms that have the properties above have been improved, and some other algorithms could beat many fair opponents, but now there is not an algorithm have these two merits at the same time.In this paper, firstly, we analyzed the behavior of agents that incrementally adapt their strategy through gradient ascent on expected payoff, in the simple setting of two-player, two-action, iterated general-sum games. Secondly, we proposed a new deduction according to the surprising result concluded by S. Singh. Then an improved algorithm based on our deduction is proposed, which could satisfy these two merits mentioned above. Many experiment results obtained in Game Pennies-Matching and other game against some other algorithms will be given. We also proposed some theory that multi-agent learning algorithm would cover probably in the future.