Analyzing sensor states and internal states in the tartarus problem with tree state machines

Abstract

The Tartarus problem is a box pushing task in a grid world environment. It is one of difficult problems for purely reactive agents to solve, and thus a memory-based control architecture is required. This paper presents a novel control structure, called tree state machine, which has an evolving tree structure for sensorimotor mapping and also encodes internal states. As a result, the evolutionary computation on tree state machines can quantify internal states and sensor states needed for the problem. Tree state machines with a dynamic feature of sensor states are demonstrated and compared with finite state machines and GP-automata. It is shown that both sensor states and memory states are important factors to influence the behavior performance of an agent. © Springer-Verlag 2004.