This study aims to enable a high-precision landing on the moon using small unmanned probes in space. To achieve this goal, matching images taken by a space probe with map images of the lunar surface is being considered. However, camera anomalies caused by radiation are an obstacle, to high-precision image matching. To address this problem, random sample consensus (RANSAC) is used as an outlier removal method. However, because RANSAC has the problem that it searches randomly, a method that combines it with reinforcement learning is effective. Various types of reinforcement learning methods exist, such as softmax method and the upper confidence bound (UCB) policy. However, these have not been comprehensively evaluated. We focus on the UCB policy, which is expected to provide stable data in a deterministic policy. In this study, we extend the deterministic UCB policy to a probabilistic policy, experimental evaluation of these reinforcement learning high-precision matching is conducted.