A Bioinspired Approach for Mental Emotional State Perception towards Social Awareness in Robotics
Jordan J. Bird, Diego R. Faria · Journal of robotics & autonomous systems · 2019
This preliminary study explores a new approach to EEG data classification by using the concept of evolutionary algorithms to perform attribute selection, as well as optimise a neural network for data classification in mental communication for robotics. EEG brainwave data is recorded from a preliminary set of subjects via the TP9, AF7, AF8, and TP10 electrodes used by the EEG headband, and 2550 statistical temporal features are extracted as dimensions of data. Nature inspired evolutionary algorithms select attributes before an evolutionary algorithm optimizes a neural network topology. A Long Short-Term Neural Network is also trained to perform deep learning on the data. Promising results show that the evolutionary optimised neural net scores 96.11% accuracy and the LSTM achieves 96.86%. The evolutionary neural network, although lacking in 0.75 accuracy points, has a training time far more optimal than the LSTM, at less than 25% of the required resource usage.