Stochastic synapse reinforcement learning (SSRL)

摘要

Over the past several decades, reinforcement learning has emerged as one of the major paradigms in machine learning because it allows an agent to learn through interaction with its environment, so long as there is some mechanism by which the agent can gain evaluative feedback on the effects of its actions. However, there are still many open questions as to the most appropriate reinforcement learning approach, particularly for difficult problems such as those dealing with delayed reward, unknown reward structures, continuous state and/or action spaces, perceptual aliasing, and/or environmental change. Here we present a new learning algorithm for these types of difficult problems. It combines the eligibility traces approach to reinforcement learning with artificial neural networks for generalization and pushes the idea of stochastic computations down to the level of the synapse. A proof-of-concept experiment in the domain of robotics demonstrates that the approach has promise.