A New Hybrid Model for Associative Reinforcement Learning

In this paper, a new model, addressing the Associative Reinforcement Learning (ARL) problem, based on learning automata and self organizing map is proposed. The model consists of two layers. The First layer comprised of a SOM which is utilized to quantize the state (context) space and the second layer contains of a team of learning automata which is used to select an optimal action in each state of the environment. First layer is mapped to the second layer via an associative function. In other words, each learning automaton is in correspondence with only one neuron of the self organizing map. In order to show the performance of the proposed method, it has been applied successfully to classification applications on Iris, Ecoli, and Yeast data sets, as examples of ARL task. The results of experiments show that the proposed method is reached the accuracy near to or even higher than the highest reported accuracy. The results obtained for Ecoli and Yeast data sets indicate that the method is able to classify in relatively high dimensional context space and high number of classes.