Abstract
In this paper, a forgettable rational Prot Sharing reinforcement learning method, which can
learn much faster than conventional method without no additional parameters, is proposed. Furthermore,
this method can suit for environment changes very quickly. Reinforcement Learning (RL) is one of the
eective unsupervised learning techniques. However, the most of RL method do not immediately present
an appropriate solution, even though the size of problem is small. Prot Sharing (PS) is a novel method
to solving this diculty. It learn better solution much faster, rather than learning optimum solutions
denitely . However, its learning speed becomes quite slow when a learner try to learn long term tasks.
There are several approaches to accelerate its learning speeds, however, these method relatively converge
to inappropriate solutions. The proposed method lets learner to forget a particular experiences that make
its policy worse in order to improve its policy. Two numerical examples are demonstrated to show its
eectiveness.
