Reinforcement Learning via Recurrent Convolutional Neural Networks
This work addresses the challenge of partially observable environments in reinforcement learning for robotics, offering an incremental improvement over existing methods.
The paper tackled the problem of model-free reinforcement learning ignoring task structure by introducing Recurrent Convolutional Neural Networks (RCNNs) to better exploit inherent structure, resulting in reduced replanning costs and learning accurate MDP models for near-optimal policies in a robot planning simulation.
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 achieve considerable performance, they often ignore the structure of task. We present a natural representation of to Reinforcement Learning (RL) problems using Recurrent Convolutional Neural Networks (RCNNs), to better exploit this inherent structure. We define 3 such RCNNs, whose forward passes execute an efficient Value Iteration, propagate beliefs of state in partially observable environments, and choose optimal actions respectively. Backpropagating gradients through 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 replanning, learn accurate MDP models, and finally re-plan with learnt models to achieve near-optimal policies.