Deep Reinforcement Learning algorithm for Pattern Recognition

Reinforcement learning

Total structure

• The electromyography(EMG) signals were measured by Myo Arm band.
• Time domain features were extracted by various feature extraion algorithm.
• While agent is training, human’s feedback which is SSVEP were used as reward.
• Data sets were splited into training, test set, and 5-fold validation were used.

• Using SSVEP estimation, human can give the right answer to reinforcement learning agent
• Each Motion was mapped into phillips Hue smart light bulb.
• Human feedback put into replay memory.

Result

• RL agent predicts the human arm motion, and control the light bulb.
• If the control was wrong, SSVEP result will be used as a human feedback to train again the agent.

Deep Reinforcement Learning algorithm for Pattern Recognition

Introduction

• Conventional methods need the label(ground truth) to train the supervisd learning model. By utilizing the SSVEP(Steady State Visually Evoked Potential), it is possible to get a  human feedback as  a reward.
• By using human’s feedback, human arm motion pattern recognition accuracy will be increased.
• Reinforcement learning agent chooses the optimal features for human motion pattern recognition, and classifies the human motion patterns.
• The project aims to develop and distribute an open source Reinforcement learning Pattern recognition project and associated developer guide.
• It is working base on Myo-python library, Python 3.6, PyTorch .