Abstract
A reinforcement learning based on neural networks is able to treat continuous values which describe any statuses in the real world, and is thought as useful learning algorithm for object control. The feed-forward neural network, one of popular neural networks, needs to construct appropriate network configuration to accomplish learning. The reinforcement learning is applied to controlling objects of which control rules are unknown, thus it is difficult to predict network configuration for efficient learning performance. We have proposed the reinforcement learning algorithm constructed from neural network which has mechanism adding or removing neurons. In this report, we show the results that are evaluation of the proposed reinforcement learning algorithm by the mountain-car task as a new benchmark task.
