EFFICIENCY ANALYSIS OF THE TIME-SHUFFLING METHOD FOR THE EVOLUTION OF AGENT BEHAVIOR
We have analyzed the effectiveness and the efficiency of a time-shuffling method applied to an evolutionary algorithm scheme in order to optimize the behavior of autonomous agents in a multi-agent system. The multi-agent system is modeled as cellular automata (CA) because of the inherent parallelism of the model, which suits well the requirements of a system of autonomous moving agents with a local view. The task of the agents is the all-to-all communication, i.e., all agents shall communicate their initially mutually exclusive information to all other agents. The agents' uniform behavior is defined by a finite-state machine, which is evolved by a genetic algorithm (GA). 20 different initial two-dimensional environments were defined as a training set, 10 of them with border, 10 with cyclic wrap-around. The state machine was evolved (1) directly by a GA for all 20 environments, and (2) indirectly by two separate GAs for the 10 environments with border and the 10 environments with wrap-around, with a subsequent time-shuffling technique in order to integrate the good abilities from both of the separately evolved state machines. The time-shuffling technique alternates two state machines periodically. The results show that time-shuffling two separately evolved state machines is effective and much more efficient than the direct application of the GA.