Forearm movements classification via genetic algorithm optimized fuzzy reinforcement learning: An approach towards prosthesis design

This work proposes a Genetic Algorithm optimization based Fuzzy Reinforcement Learning (GAFRL) classifier for six forearm movements. Electromyogram (EMG) signal for movements are acquired through Biopac 100c System at asampling frequency of 1000 Hz for 20 subjects. In the first step, features are extracted through Hilbert Huang transform (HHT) and Fuzzy lattices formation. Both HHT and Fuzzy lattices formation techniques have been used to handle non-stationary and non-linear behaviour of the EMG signals. In HHT, signals are decomposed to find Intrinsic Mode Functions (IMFs) and features for these IMFs are extracted. In Fuzzy lattices formation technique, Kinetic Energy features are computed via Schrödinger equation for these lattices. Thereafter, GA optimized Fuzzy Reinforcement Learning classifier (self-learning) has been employed to identify forearm movements based on features selected via Principal Component Analysis (PCA). The proposed classifier learns without knowledge of targets and learn on its own through a reward/punishment policy. The average success rate (training) of Fuzzy Reinforcement Learning (FRL) and GAFRL is 93.491 % and 95.475 % respectively. The FRL classifier achieves an Accuracy, Sensitivity, and Specificity of 93.4 %, 94.6 % and 92.25 % respectively and classifier GAFRL achieves an Accuracy, Sensitivity, and Specificity of 95.8 %, 97.8 % and 93.8 % respectively. These results are the average values for six forearm movements (hand close, open, supination, pronation, flexion and extension). The proposed technique is an adaptive self-learning one, it could be useful in designing prosthetic arms for movement controls.