Deep Reinforcement Learning Apply in Electromyography Data Classification

Abstract

In this paper, we try a novel approach to detect human movement intentions based on electromyography(EMG) signals. We use the Convolutional neural network(CNN) to extract EMG features automatically, then use dueling deep Q-learning, a reinforcement learning technique, to learn a classification policy which with an ability to select most helpful subset features and filter the irrelevant or redundant features from the deep learning features. We show that the deep learning method outperforms the multi-layer perceptron in the several subjects EMG data classification situation and the reinforcement networks can use less features to reach a relatively high classification precision.