Neural network generation of muscle stimulation patterns for control of arm movements

The authors present a new approach to the control of point-to-point, single joint arm movements by an artificial neural network (ANN) controller. The ANN controller was used to learn and store the optimal patterns of muscle stimulation for a range of single joint movements. These stimulation patterns were obtained from an optimal control strategy that minimizes muscle activation or muscular effort. Feedforward, recurrent feedback, and time delay topologies of neural networks were considered for this application. The choice of a network structure was based on the learning performance and ability to generalize a learned muscle simulation pattern to novel movements. A comparison showed that the feedforward network combined with recurrent feedback and input time delays can most effectively capture the optimal temporal profiles of muscle stimulation. This neural network controller further demonstrated remarkable ability to generalize the learned optimal control to a class of scaled movements. The authors also evaluated open-loop performance of movement control by the ANN with a nonlinear muscle/joint model. The trained neural network controller reproduced the range of scaled optimal movements well, though sometimes with terminal position errors. This study showed that neural networks were promising as an open-loop pattern generator for muscle stimulation signals in movement restoration by functional electrical stimulation