Abstract:
The automation of tasks such as environmental monitoring, toxin detection, and mineral resource identification requires
artificial agents with perceptual discrimination capabilities to identify the predominant features in environments much larger than their
sensing range. The key challenge is developing collective decision-making methods that allow agents to predict a global perspective of
the environment from local observations. Our research explores the effectiveness of collective decision-making for binary perceptual
discrimination tasks, controlled by an artificial neural network synthesised using evolutionary computation techniques. We focus
on strategies that generalised better to environments with patchy, clustered feature distribution. We investigate three communication
strategies - close-neighbour, rand-neighbour, and far-neighbour- in which robots exchange opinions about the dominant colour of the
environment based on the distance between sender and receiver robots. The results show that the rand-neighbour strategy significantly
improves performance, particularly in unseen patchy patterns. The extensive analysis of the communication dynamics among the
robots indicates that the effectiveness of rand-neighbour strategy is attributed to its efficient circulation of opinions among both close
and distant robots. Our findings support the hypothesis that primordial communication between one receiver robot and a randomly
chosen emitter robot is sufficient to develop an effective collective decision-making strategy for swarm of robots engages in perceptual
discrimination tasks.