Mitsuo KAWATO
Optimization and Learning
in Neural Networks
for Formation and Control of
Coordinated Movement
Abstract:We proposed that the trajectory followed by human subject arms tended
to minimize the time integral of the square of the rate of change of torque.
Based on this computational model, the cascade neural network model, which
utilizes a forward model of a controlled object, reproduced Fitts's law of
speed-accuracy tradeoff as well as various invariant features of path and
velocity profiles of multi-joint arm movement. For supervised motor learning,
conversion of the error signal calculated in the task space into that of
the motor command space is most essential and difficult. We proposed a
feedback-error-learning approach in which the feedback motor command is
used as an error signal to train an inverse model of the controlled object,
which then generates a feedforward motor command. Here, we propose a
unified neural network model which integrates the two previous models. In
this model, for very skilled movements relaxation computation is conducted
using both the forward and inverse models of the controlled object, while
only the inverse model acquired by the feedback-error-learning is utilized for
relatively difficult or less skilled movements.