Evolution of neural architecture fitting environmental dynamics

Temporal and sequential information is essential to any agent continually interacting with its environment. In this paper, we test whether it is possible to evolve a recurrent neural network controller to match the dynamic requirement of the task. As a benchmark, we consider a sequential navigation task where the agent has to alternately visit two rewarding sites to obtain food and water after first visiting the nest. To achieve a better fitness, the agent must select relevant sensory inputs and update its working memory to realize a non-Markovian sequential behavior in which the preceding state alone does not determine the next action. We compare the performance of a feed-forward and recurrent neural control architectures in different environment settings and analyze the neural mechanisms and environment features exploited by the agents to achieve their goal. Simulation and experimental results using the Cyber Rodent robot show that a modular architecture with a locally excitatory recurrent layer outperformed the general recurrent controller.