Analysis of Ultrastability in Small Dynamical Recurrent Neural Networks

This paper reconsiders Ashby’s framework of adaptation within the context of dynamical neural networks. Agents are evolved to behave as an ultrastable dynamical system, without imposing a priori the nature of the behavior-changing mechanisms, or the strategy to explore the space of possible dynamics in the system. We analyze the resulting networks using dynamical systems theory for some of the simplest conditions. The picture that emerges from our analysis generalizes the idea of ultrastable mechanisms.