Abstract
This paper describes a novel dynamical neural network model for learning and generating object manipulation behavior. The network learns to predict not only the mean of the next input state, but also its timedependent variance. The training method is based on maximum likelihood estimation by using the gradient descent method, and the likelihood function is expressed as a function of the estimated variance. Regarding the model evaluation, it was shown that a humanoid robot with the proposed network can learn multiple behavior of object manipulation and adaptively select trained behavior in accordance with the environmental changes.
