Assessment of evolutionary processes

Abstract

Evolutionary process has become a popular design method for experimenting and automatically synthesizing intelligent controllers for autonomous robots. Such controllers are automatically created using different evolutionary methods without direct programming or in-depth human knowledge of the design. Multi-agent systems and collective behaviors based on swarm intelligence observed in nature are generally ideal candidates for automatic controller design using evolutionary processes. Although the evolutionary process can provide great insight into possible solutions and is a reasonable tool for such experiments, it may not be the most efficient and ideal design tool for every experiment. In this paper, we setup experiments on self-organization of a multi-agent system and evolve three different controllers. We then compare the design effort and results of the three evolved controllers with a traditionally designed controller that performs the same task. We show that the traditionally designed controller outperforms the best case evolutionary-based controller by 10% when measured in overall median fitness level. We also show that the traditionally designed controller is the most reliable as it never violates the design rules. Many studies have been performed in comparison of different evolutionary techniques, but to the best of our knowledge none of them focus on the study of design effort and suitability of such approaches. The evolutionary process is a reasonable design tool only for problems that are too difficult or complicated to be addressed using traditional design methods.