Reinforcement Learning via Recurrent Convolutional Neural Networks

摘要

Deep Reinforcement Learning has enabled the learning of policies for complex tasks in partially observable environments, without explicitly learning the underlying model of the tasks. While such model-free methods do achieve considerable performance, they often ignore the structure of task. We present a more natural representation of the solutions to Reinforcement Learning (RL) problems, within 3 Recurrent Convolutional Neural Network (RCNN) architectures to better exploit this inherent structure. The forward passes of each RCNN execute an efficient Value Iteration, propagate beliefs of state in partially observable environments, and choose optimal actions respectively. Applying back-propagation to these RCNNs allows the system to explicitly learn the Transition Model and Reward Function associated with the underlying MDP, serving as an elegant alternative to classical model-based RL. We evaluate the proposed algorithms in simulation, considering a robot planning problem. We demonstrate the capability of our framework to reduce the cost of re-planning, learn accurate MDP models, and finally re-plan with learned models to achieve near-optimal policies.