Dynamics from patterns: creating neural controllers with SENMP

Abstract

In this paper we show how simple laterally interacting computational entities, i.e. neurons, can be guided by a selectionist process into spatial patterns that show interesting and purposeful dynamics with regard to a particular utility measure. In other words, if a suitable population of laterally interacting mobile entities exist, it is possible to gradually arrange the entities into a spatial pattern that exhibits the desired dynamics. In this paper, the selectionist process is implemented with the stochastic evolutionary neuron migration process (SENMP) and approach is used to evolve dynamic recurrent neural networks (DRNNs) for controlling complex dynamic systems such as autonomous mobile robots, for example. The feasibility and advantages of the approach are demonstrated by evolving neural controllers for solving a non-Markovian double pole balancing problem. In addition, we have earlier used SENMP to evolve navigation behaviors for mobile robots in complex simulated and real environments.