View article

This paper presents a study investigating the generalization characteristics of two neuro-controllers underpinning decision-making mechanisms in a swarm of robots engaged in a collective perception task. The neuro-controllers are both designed using evolutionary computation to operate in a randomly distributed cues environment, but under different conditions for what concerns the length of the communication range characterising the robots interactions. For one neurocontroller, the communication range during the design phase is 30 cm, while for the other is 50 cm. The aim of the study is to explore the robustness and the limitations of the two distinct neuro-controllers across a range of different conditions and to establish the optimal bounds on the swarm communication range for this collective perception task. To examine the performance of the two neuro-controllers we conduct a series of post-evaluations in 45 distinct environments, given by nine different distributions of the perceptual cues, and five different communication ranges (ie, 10, 20, 30, 40, and 50 cm). The results demonstrate that the neurocontroller evolved with a swarm communication range of 30 cm generalizes better and exceeds the performance of the other neuro-controller evolved with 50 cm communication range in a vast majority of the postevaluation conditions. However, the swarm performance degrades in conditions with patchily distributed perceptual cues and/or very short communication range.