Shaped prisoner's dilemma automata

Abstract

Previous research has shown that game playing agents evolved with different representations behave differently. Finite state automata are one of the common representations used to encode agents to play the iterated prisoner's dilemma. They comprise a large search space with many possible distinct behaviors. In this paper we explore the consequences of evolving agents within subsets of this space by limiting their shape. A shape for a finite state automaton is a restriction on what transitions are permitted out of each state in the automaton. Eight shapes for agents that play iterated prisoner's dilemma, including a baseline shape with all possible transitions, are tested with an evolutionary algorithm by enforcing the shape during the co-evolution of agents. All eight shapes yield distinct distributions of behaviors in the evolved agents. Some of the observed types of play are entirely novel.