A new hillclimber for classifier systems

Abstract

Multi-state artificial environments such as mazes represent a class of tasks that can be solved by many different multi-step methods. When different rewards are available in different places of the maze, a problem solver is required to evaluate different positions effectively and remembers the best one. A new hillclimbing strategy for the Michigan style classifier system is suggested which is able to find the shortest path and discarding sub-optimal solutions. Knowledge reuse is also shown to be possible. 1 Introduction Classifier Systems (CSs) have been used to study complex the emergent behaviour of artificial creatures in simulated environments [2], commonly using robots both real and simulated [1, 7]. The seminal work by Holland and Reitman [3] used a simulated creature in the context of a one dimensional maze. Others [4, 6] have used a multi-state environment where a robot is required to transfer itself from one state to another to test or demonstrate various behaviours of CSs. M...