Homeostatic plasticity improves continuous-time recurrent neural networks as a behavioural substrate

Abstract

Homeostatic plasticity is applied to continuous-time recurrent neural networks. It is observed to make networks more sensitive, improve signal propagation and increase the likelihood of autonomous oscillations. Evolutionary experiments with a simulated robot show that in some circumstances homeostatic plasticity improves evolvability of good control networks, but in others it makes good controllers less easy to evolve.