Realization of reinforcement learning using multi-winners KFM associative memory

In this paper, we propose a multi-winners Kohonen feature map (KFM) associative memory, and apply it to reinforcement learning. In the proposed model, the patterns are trained by the successive learning algorithm of the conventional KFM associative memory. The proposed model has two kinds of recall methods, and one of them is selected based on whether or not the input pattern is the trained pattern. In one of the recall method, the output of the input/output layer is calculated as the weighted sum of the connection weights of the fired neuron in the map layer according to their internal states. In the other one method, one of the weight-fixed neurons are selected in the map layer, and the output of the input/output layer is determined based on the connection weights of the neuron. In the reinforcement learning, the proposed model can select the trained corresponding action if the known environment is given. Moreover, it can select appropriate action based on the trained similar situation even if the unknown environment is given.