Towards the light — Comparing evolved neural network controllers and Finite State Machine controllers

Abstract

In this paper, we compare two different evolvable controller models based on their performance for a simple robotic problem, where a robot has to find a light source using two luminance sensors. The first controller is a fully meshed artificial neural network. Though neural networks are the most common type of controllers used in evolutionary robotics, validating and understanding the resulting neural network is problematic. In order to overcome this problem, we implement also an evolvable Mealy machine, which is a specific Finite State Machine. We show that both controllers can be evolved with evolutionary algorithms to find a light source placed outside the sensor range of the robots, but the evolved neural network controller shows better performance in speed and success probability, while the internal structure of the evolved Mealy machine is more comprehensible.