Automatic Synthesis of Working Memory Neural Networks with Neuroevolution Methods

Abstract

Evolutionary Robotics is a research field focused on autonomous design of robots based on evolutionary algo-rithms. In this field, neuroevolution methods aim in par-ticular at designing both structure and parameters of neu-ral networks that make a robot exhibit a desired behavior. While such methods have shown their efficiency to gener-ate reactive behaviors, they hardly scale to more cognitive behaviors. One of the reasons of such a limitation might be in the properties of the encoding, i.e. the neural net-work representation explored by the genetic operators. This work considers EvoNeuro encoding, an encoding directly inspired from computational neuroscience [1] and tests its efficiency on a working memory task, namely the AX-CPT task. Neural networks able to solve this task are generated and compared to neural networks evolved with a simpler direct encoding. The task is solved in both cases, but while direct encodings tend to generate networks whose structure is adapted to a particular instance of AX-CPT, networks generated with EvoNeuro encoding are more versatile and can adapt to the new task through a simple parameter op-timization. Such versatile neural network encoding might facilitate the evolution of robot controllers for tasks requir-ing a working memory. KEYWORDS evolutionary algorithms; neural networks; computational neuroscience; working memory; neuroevolution. 1