A learning algorithm for continually running fully recurrent neural networks
Neural Computation
Hi-index | 0.00 |
A computational model in the framework of reinforcement learning, called Cooperative Motor Learning (CML) model, is proposed to realize cerebellar control of balance and locomotion. In the CML model, cerebellum is a parallel pathway that the vermis and the flocculonodular lobe play a role of reflex actions executor, and that the intermediate zone of the cerebellum participates in initiating voluntary actions. During the training phase, the cerebral cortex provides the predictive error through climbing fiber to modulate the concurrently activated synapses between parallel fibers and purkinje cells. Meanwhile, the intermediate zone of the cerebellum computes the temporal difference (TD) error as a training signal for the cerebral cortex. In the simulation experiment for the balance of double inverted pendulum on a cart, a well-trained CML model can smoothly push the pendulum into the equilibrium position.