Automatic Creation of an Efficient Multi-Agent Architecture Using Genetic Programming with Architecture-Altering Operations
The MIT Press eBooks · 1996
Previous work in multi-agent systems has required the human designer to make up-front decisions about the multi-agent architecture, including the number of agents to employ and the specific tasks to be performed by each agent. This paper describes the automatic evolution of these decisions during a run of genetic programming using architecture-altering operations. Genetic programming is extended to the discovery of multi-agent solutions for a central-place foraging problem for an ant colony. In this problem each individual ant is controlled by a set of agents, where agent is used in the sense of Minsky's Society of Mind. We describe the simultaneous evolution of the number of agents needed to solve the problem and the work performing steps of each agent. Genetic programming was able to evolve time-efficient solutions to this problem by distributing the functions and terminals across successively more agents in such a way as to reduce the maximum number of functions executed per agent. The other source of time-efficiency in the evolved solution was the cooperation that emerged among the ants in the ant colony.