Improvement for action strategy learning in classification task using classification probalilities

In this paper, we address the autonomous evidence accumulation when a system classifies an object into one of predetermined categories. We propose a reinforcement learning system that effectively selects actions to speed up the classification process. The proposed system accelerates its learning using classification probabilities calculated by a classification system. We conducted three binary classification experiments to evaluate the learning speed and correctness of the proposed system. In the first experiment, we examined a random action selection strategy that does not learn its selection parameters while accumulating evidence. In the second experiment, we examined Paletta's reinforcement learning system that observes the state of the object and learns action selection strategy. In the third experiment, we examined the proposed system that observes both the object state and the classification probability. The proposed system showed the fastest learning.