Real Time Multiagent Decision Making by Simulated Annealing

Abstract

Computer data. In this chapter, SA algorithm is employed to solve real time multiagent decision making problem. Compared with exact method this chapter's empirical evidences show that (1) this method is almost optimal with a small fraction of the time that VE takes to compute the policy of the same coordination problem; (2) the running time of SA grows linearly with the increasing number of neighbors per agent;(3) it is an anytime algorithm which return result at any time. For above reasons, it is believed that SA is a feasible approach for action selection in large complex cooperative autonomous systems. As future research, an appropriate setting of the acceptable probability will be figured out, especially the decay rate in SA. Some recent work shows that neural network algorithm can produce a good decay rate for larger problems. Such techniques may be employed to solve multiagent decision making problem. Furthermore, whether reinforcement learning algorithms can be applied to automatically learn the payoff in each value rule is to be investigated.