Using Multi-Agent Systems for Learning optimal Policies for complex Problems
Abstract
The automatic computation of an optimal solution for a complex problem is a challenging task if no additional knowledge is available. For bounded sized problems there are universally applicable algorithms (e.g. genetic algorithms, branch and bound, reinforcement learning). The disadvantage of these algorithms is their high computational complexity so that real world problems can only be solved efficiently, if the search space is reduced dramatically. In this paper we present an approach that enables the automatic computation of the parameter dependencies of a complex problem without any additional information. The basic idea is to apply reinforcement learning and to incrementally acquire knowledge about the implicit parameters dependencies. Based on the obtained data an optimal strategy is learned. For speeding up the learning process a multi-agent architecture is applied, that supports the simultaneous analysis of alternative strategies. We prove the advantages of our approach by successfully learning a control strategy for a model helicopter.