Comparative Analysis of Temporal Difference Learning Methods to Learn General Value Functions of Lower-Limb Signals.

Abstract

Millions of people in the United States suffer from paralysis, resulting in significant deficits in motor function. Restricted mobility due to these deficits and the lack of adaptive rehabilitative solutions make traversing complex and challenging terrains unsafe. Exoskeletons offer a promising solution, but their effectiveness could be greatly enhanced by incorporating reinforcement learning algorithms for real-time adaptation to changing environments and the user's unique gait biomechanics. This study explored different temporal difference learning methods for predicting signals recorded from sensors on the lower-limbs, including muscle activation from electromyography, underfoot pressure, and joint angles from goniometers. Specifically, the performance of the temporal difference learning methods TD $(\lambda)$, TOTD, and SwiftTD to quickly and accurately predict these signals was examined. From initial findings, SwiftTD generally converged faster, while TOTD typically achieved lower convergence errors. These outcomes varied depending on the specific signal that was being predicted, highlighting the need for careful consideration of algorithm choice depending on the signal, accuracy, and speed. The results, therefore, support the informed selection of specific algorithms for providing predictive knowledge to adaptive, machine learning-controlled assistive rehabilitative technologies. These findings will enable the selection of appropriate predictive algorithms, leading to the development of better exoskeletons and other assistive devices to enhance the mobility and quality of life of individuals with motor paralysis.