Profit Sharing That Can Learn Deterministic Policy for POMDPs Environments by Kohonen Feature Map Associative Memory

In this paper, we propose a Profit Sharing that can learn deterministic policy for POMDPs environments by Kohonen Feature Map (KFM) associative memory. The proposed method is based on the conventional Profit Sharing that can learn deterministic policy for POMDPs environments which can obtain the deterministic policy by using the history of observations and the variable-sized Kohonen Feature Map Probabilistic Associative Memory (for short, KFM associative memory). In the conventional Profit Sharing that can learn deterministic policy for POMDPs environments, robustness for noise does not guaranteed. In the proposed method, rules which are obtained in the Profit Sharing that can learn deterministic policy for POMDPs environments are trained by the KFM associative memory and appropriate action selection is realized in noisy environment. We carried out a series of computer experiments, and confirmed that the proposed method can learn rules for deterministic action selection as similar as the conventional Profit Sharing that can learn deterministic policy for POMDPs environments, and appropriate action selection is realized in noisy environment.