'Psychoanalysis' of a Minimal Agent

The Secretary problem is studied with minimal cognitive agents, being a problem that needs memory and judgment. A sequence of values, drawn from an unknown range, is presented; the agent has only one chance to pick a single value as they are presented, and should try to maximize the value chosen. In extension of previous work (Tuci et al. 2002), Continuous Time Recurrent Neural Networks (CTRNN) are evolved to solve the problem, and then their strategies are analyzed by relating mechanisms to behavior. Strategies similar to the known optimal strategy are observed, and it is noted that significantly different strategies can be generated by very different mechanisms that perform equally well.