Macroaction Synthesis for Agent System

We present methods to synthesize macroactions for agent systems and the methods are combined with SOS algorithm that learns rules for agent´s behavior using reinforcement learning and evolutionary computation. To acquire useful macroactions, our methods use some kinds of numerical values evaluated in performing SOS algorithm, e.g., fitness values of actions or the number of transitions between rules. New macroactions generated by our methods are fed back to SOS algorithm for learning rules. By repeating macroaction synthesis and learning rules alternately, rules for agent´s behavior are acquired. The methods shown here have been implemented and some experimental results have been shown.