The standard approach to creating a machine learning based controller is to provide users with a number of gestures that they need to make; record multiple instances of each gesture using specific sensors; extract the relevant sensor data and pass it through a supervised learning algorithm until the algorithm can successfully identify the gestures; map each gesture to a control signal that performs a desired outcome. This approach is both inflexible and time consuming.
The primary contribution of this research was to investigate a new approach to putting artificial intelligence into wearable human-machine interfaces by creating a Generic, Self-Improving Controller. It was shown to learn two user-defined static gestures with an accuracy of 100% in less than 10 samples per gesture; three in less than 20 samples per gesture; and four in less than 35 samples per gesture. Pre-defined dynamic gestures were more difficult to learn. It learnt two with an accuracy of 90% in less than 6,000 samples per gesture; and four with an accuracy of 70% after 50,000 samples per gesture.
The research has resulted in a number of additional contributions:
• The creation of a source-independent hardware data capture, processing, fusion and storage tool for standardising the capture and storage of historical copies of data captured from multiple different sensors.
• An improved Attitude and Heading Reference System (AHRS) algorithm for calculating orientation quaternions that is five orders of magnitude more precise.
• The reformulation of the regularised TD learning algorithm; the reformulation of the TD learning algorithm applied the artificial neural network back-propagation algorithm; and the combination of the reformulations into a new, regularised TD learning algorithm applied to the artificial neural network back-propagation algorithm.
• The creation of a Generic, Self-Improving Predictor that can use different learning algorithms and a Flexible Artificial Neural Network.