Testing a Protocol for Characterizing Game Playing Agents Trained via Evolution on a New Game

A large series of studies on evolving agents to play mathematical games has demonstrated that many factors can significantly impact which agents arise, when those agents arise during evolution, and how robust they are in their play against other agents. Some or all of these factors have been shown to be relevant in the iterated prisoner's dilemma, the snowdrift game, and a fairly complex game called divide-the-dollar. This study demonstrates the impact or representation and agent resource allocation for a new game called coordination prisoner's dilemma. This paper demonstrates protocols from a recently published book for analysis of agent behavior and extends the work to another game, the first three-move game so treated. A new representation for agents playing mathematical games is introduced, a linear genetic programming register machine. New metrics for agent behavior including total exploitation, strategic variability, and action entropy are introduced. It is found that varying the representation and resource levels within a representation changes the types of game playing agents produced by evolution for coordination prisoner's dilemma.