Detecting evolutionary bias in the decision-making of robotic swarms

This paper presents a simple and easy to use, yet powerful method for identifying evolutionary bias in neuro-controllers, based on classical Shannon entropy in one specific domain of swarm robotics. We demonstrate its application for assessing the quality of one neural network-based decision-making controller for swarm opinion formation, evolved for the site selection task. The preliminary results, based on simulated swarm behaviour recorded in three experimental conditions and nine trials overall, reveal the benefits and the consistency of the applied measure and its ability to discern various opinion formation trends. Such a tool could help automate the process of bench-marking controllers and assess their quality with respect to evolutionary bias, and thus minimize designers effort, while at the same time provide better models performing consistently across conditions.