Recognition of Hand Gestures Based on EMG Signals with Deep and Double-Deep Q-Networks

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Machine Learning

Background:

• Hand gesture recognition (HGR) is crucial for human-machine interfaces.
• Electromyography (EMG) signals are increasingly used for HGR.
• Supervised learning dominates HGR, but reinforcement learning (RL) offers potential advantages.

Purpose of the Study:

• To develop and evaluate a user-specific HGR system using RL.
• To compare Deep Q-Network (DQN) and Double-Deep Q-Network (Double-DQN) for EMG signal classification.
• To assess the impact of Long Short-Term Memory (LSTM) integration on RL-based HGR performance.

Main Methods:

• Implemented RL agents using DQN and Double-DQN algorithms.
• Employed feed-forward artificial neural networks (ANNs) for agent policy representation.
• Tested ANN with and without an added LSTM layer for performance comparison.
• Utilized the public EMG-EPN-612 dataset for training, validation, and testing.

Main Results:

• The DQN model without LSTM achieved the highest classification accuracy (90.37%±10.7%).
• Recognition accuracy for the best model reached 82.52%±10.9%.
• RL methods demonstrated competitive performance in EMG-based HGR tasks.

Conclusions:

• Reinforcement learning, specifically DQN, shows significant promise for EMG-based hand gesture recognition.
• The proposed user-specific RL approach offers an effective alternative to traditional supervised methods.
• Further research into RL for EMG signal processing can advance human-machine interface development.