[1] Graduate School of Human Arts and Sciences, Kurashiki University of Science and the Art
[2] Department of Physical Therapy, Takarazuka University of Medical and Health Care
[3] Department of Rehabilitation Medicine, Ishikawa General Hospital
The primary objective of this study is to conduct computer simulations for analyzing the variations in a human body while in the standing position with an open leg by employing deep Q-learning. To achieve the said objective, an equation representing angular motion was developed for indicating the slant angle of the body to the vertical. Additionally, a neural network comprising an input layer, three middle layers, and an output layer was constructed. The input layer includes six nodes wherein data corresponding to the angle, angular velocity, angular acceleration, active torque, total torque, and life time of active torque are respectively submitted. Each of the three middle layers comprises thirty-six nodes, and each node features the ReLU-function as the activation function. The output layer consists of six nodes whose output values determine the action of an agent. Upon completion of the entire learning process by an agent (a virtual human) through the neural network and reinforcement learning, the neural network’s parameter weights acquired the most appropriate values. By incorporating such parameters, computer simulations for the angular motion of an agent could be conducted. Furthermore, in the proposed study, results obtained from simulations are compared with corresponding measurement values, and the future scope of this study is discussed.
Keyword: Reinforcement learning, Deep reinforcement learning, Deep Q-learning, Open leg standing upset problem, Computer simulation