Performance Evaluation of Neural Architectures for Sequential Tasks

摘要

This paper considers a sequential task where the agent has to alternatively visit two rewarding sites to obtain food and water after first visiting the nest. To achieve a better fitness, the agent must have a working memory to reach the target position, must ignore irrelevant sensory inputs, and at a higher level, it has to deal with the non-Markovian order of sequential task in which the preceding state alone does not determine the next action. We compare the performance of 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 specific architecture outperformed the general recurrent controller.