A novel self-organizing cerebellar instrumental learning algorithm based on CPN

Abstract

In this paper, a self-organizing cerebellar instrumental learning algorithm (SOCILA) is proposed for a self-balancing robot to achieve the balance control based on the counter-propagation network (CPN), which combined unsupervised and supervised learning. Due to the cerebellar contributions to instrumental learning, the instrumental learning algorithm is designed to change the weights from competitive layer to output layer of CPN, while the clustering to input state is implemented by fuzzy self-organizing map layer (from input layer to competitive layer)of CPN. In order to save the learning results, curiosity parameter is introduced, which is useful to instrumental learning and simulate the biological learning process. Simulations on balance learning of a two-wheeled self-balancing robot is given to illustrate the performance and applicability of the proposed learning scheme, and, as a result, progressive learning process of balance can be achieved using the proposed SOCILA method. Finally, in order to compare the learning performances we choose different number of rules to test the learning performance in balance learning.