ADVANCES IN EVOLUTIONARY AGENT LEARNING

This paper describes a novel methodology for software agent learning. Evolutionary Platform for Agent Learning (EPAL) creates both subtle and drastic changes to agent behavior. Subtle changes arise from learning task parameters. Drastic changes emerge from learning improved workflows containing new programming constructs and tasks. EPAL presents a general approach to agent learning, based on an extension to Genetic Programming that we developed. This approach is suitable for learning by any agent whose behavior can be represented as a workflow that can be further decomposed into building blocks, such as operators, tasks, and parameters. This paper describes a real-world problem that our agents learned to solve that is similar to problems encountered in Navy Fleet Battle Experiment-Juliet. Several sets of results are presented, showing increasing learning capabilities of our agents.