Evolutionary computation for bottom-up hypothesis generation on emotion and communication

Abstract

Through evolutionary computation, affective models may emerge autonomously in unanticipated ways. We explored whether core affect would be leveraged through communication with conspecifics (e.g. signalling danger or foraging opportunities). Genetic algorithms served to evolve recurrent neural networks controlling virtual agents in an environment with fitness-increasing food and fitness-reducing predators. Previously, neural oscillations emerged serendipitously, with higher frequencies for positive than negative stimuli, which we replicated here in the fittest agent. The setup was extended so that oscillations could be exapted for the communication between two agents. An adaptive communicative function evolved, as shown by fitness benefits relative to (1) a non-communicative reference simulation and (2) lesioning of the connections used for communication. An exaptation of neural oscillations for communication was not observed but a simpler type of communication developed than was initially expected. The agents approached each other in a periodic fashion and slightly modified these movements to approach food or avoid predators. The coupled agents, though controlled by separate networks, appeared to self-assemble into a single vibrating organism. The simulations (a) strengthen an account of core affect as an oscillatory modulation of neural-network competition, and (b) encourage further work on the exaptation of core affect for communicative purposes.