Evolving Cooperative Groups Using Shared Memory
Narendra Puppala* and Sandip Sen**
**Department of Mathematical and Computer Science, University of Tulsa 600 South College Avenue, Tulsa, OK 74104-3189, U. S. A
We present a coevolutionary approach to generating behavioral strategies for cooperating agent groups. We coevolve agent behavior with genetic algorithms (GAs), where one GA population is evolved per individual in the cooperative group. Groups are evaluated by pairing strategies from each population and the best strategy pairs are stored together in shared memory. To evaluate a strategy from one population, it is paired sequentially with strategies from the other population stored in shared memory. The maximum evaluation from all such pairings is used to evaluate the strategy. A pair is stored in shared memory if we encounter at least one pair previously stored there with a lower fitness value. We evaluate our approach using an asymmetric room painting domain with two agents – a whitewasher and a painter. Shared memory proved superior to random pairing in consistently generating optimal behavior patterns.