Neural Reinforcement Learning based Identifier for Typing Keys using Forearm EMG Signals

Abstract

This work proposes a neural reinforcement learning (NRL) identifier for accurate classification finger movements for typing tasks using forearm Electromyogram (EMG) signals. We first extract four key statistical features from the EMG signals (channel 1 and channel 2) corresponding to seven typing keys. Next, these features are fed to a reinforcement learning based k nearest neighbor neural classifier for identifying the keys using a "trial and error" approach. We use EMG data from typing tasks of ten subjects using two acquisition electrodes: channel 1 and channel 2. In the first part of our work, we attempt to classify typing keys using EMG data corresponding to one subject only. After sufficient learning, NRL classifier achieved an accuracy of 99.01% and 98.29% for channel 1 and channel 2, respectively. In second part of our work, we fed the EMG data of all the ten subjects to the NRL. The NRL is able to achieve a classification accuracy of 92.7%. We also employ a subspace ensemble nearest neighbor approximator yielding a classification accuracy of 94.3% with 5-cross fold validation and 97.1% with 3-cross fold validation. Results show the effectiveness and viability of using NRL for identifying typing movements using forearm EMG signals.