No-Fringe U-Tree: An Optimized Algorithm for Reinforcement Learning

Model-free reinforcement learning algorithms based on POMDP has been devised and adopted for many years. The complexity of the environment where the agent works determines the necessity of dealing with the whole observation space. Therefore, instance-based learning methods have been put forward. NSM, USM and U-Tree algorithms can present the whole observation space as instance chains, which are very beneficial to solve the relatively reasonable value of the action [1]. However, the method to build up the suffix tree is not of good efficiency and effect. Aiming to improve the two aspects, we come up with a new algorithm, No-Fringe U-Tree (NFU-Tree) to improve the generation of the suffix tree. The result of New York Driving experiment proves that the algorithm is more efficient and effective than the U-Tree algorithm.