The iterated prisoner's dilemma: Early experiences with learning classifier system-based simple agents

Prior research on artificial agents/agencies involves entities using specifically tailored operational strategies (e.g., for information retrieval, purchase negotiation). In some situations, however, an agent must interact with others whose strategies are initially unknown and whose interests may counter its own. In such circumstances, pre-defining effective counter-strategies could become difficult or impractical. One solution, which may be viable in certain contexts, is to create agents that self-evolve increasingly effective strategies from rudimentary beginnings, during actual deployment. Using the Iterated Prisoner's Dilemma problem as a generic agent-interaction setting, we use the Learning Classifier System paradigm to construct autonomously adapting ‘simple’ agents. A simple agent attempts to cope by maintaining an evolving but potentially perennially incomplete and imperfect knowledge base. These agents operate against specifically tailored (non-adaptive) agents. We present a preliminary suite of simulation experiments and results. The promise evidenced leads us to articulate several additional areas of interesting investigations that we are pursuing.