Shopkeeper strategies in the iterated prisoner´s dilemma

Many studies have evolved agents to play the iterated prisoner´s dilemma. This study models a different situation, called the Shopkeeper model of interaction, in which a state conditioned agent interacts with a series of other agents without resetting its internal state. This is intended to simulate the situation in which a shopkeeper interacts with a series of customers. In a majority of other studies agents either reset their internal state information before each new encounter or have relatively little internal state information. This means they cannot model situations such as being the customer that meets the shopkeeper after an obnoxious customer. We train shopkeeper prisoner´s dilemma agents against a variety of distributions of possible customers. The shopkeepers specialize their behavior to their customers but sometimes fail to discover maximally exploitative behaviors. The evolved shopkeeper agents are subject to fingerprint analysis and are shown to differ substantially from agents evolved with a round-robin fitness functions. Evaluation of the behavior of the shopkeeper agents with customers they did not encounter during evolution provides additional evidence that shopkeepers specialized to the customers, but did so incompletely for the more complex sets of customers.