Incremental evolution of neurocontrollers with a diffusion-reaction mechanism of neuromodulators

Abstract

Recently, the evolutionary robotics (ER) approach has attracted attention in the field of robotics and artificial life, since it can automatically synthesize controllers by taking the embodiment and the interaction dynamics with the environment into account. However, the ER approach still has several issues to be solved. One is that it becomes significantly difficult to evolve the whole controller in one go (1-shot evolution), as task complexity and/or environment interaction dynamics increases. This is called the bootstrap problem. It is necessary to develop a method that can efficiently increase a controllers' ability without impairing the previously obtained ones. We propose a method for incremental evolution of neurocontrollers by introducing a receptor-ligand concept, the polymorphic property of neural circuits through diffusion-reaction mechanism of chemical substances, so-called neuromodulators. To investigate the effectiveness of our approach, we take a peg-pushing task, which requires an appropriate sequence of behavior to accomplish a task, as a practical example.