Challenges with reinforcement learning in prosthesis

Abstract

Reinforcement Learning has work wonders in games like Atari and AlphaZero. Recent advancement in Deep Reinforcement Learning showcase it’s ability in the active Prosthesis as well. RL is being used widely to solve problems where Learning of the Agent in its own environment is as necessary as training the model beforehand. However, model developed, and successful in the gaming environment could still need to be tuned to be effective with Real Time devices such as Prosthetic Limb and other Real-World devices. In this article, main challenges are presented which we face while working on a Model Based and Model Free Reinforcement Learning in real world environment and suggesting an approach which would work uniformly on most of the Real Time scenarios. We observed the performance and noticed that there are couple of factors which needs to be taken care of in Real Time Applications which are not much though about in games and other online applications. We also compared the algorithms such as Policy Proximal Optimization (PPO) and Deep Deterministic Policy Gradient (DDPG) vs Model Based Policy search with Gaussian Processes and found out that a mix of Model-Based and Model-Free (MBMF) performed the best individually despite of all the challenges.