Reinforcement learning algorithms for robotic navigation in dynamic environments

An issue of primary importance in the field of reinforcement learning (RL) is the tradeoff between exploration and exploitation. Balancing exploration and exploitation becomes even more crucial in an intelligent agent designed to operate in dynamic environments. Three methods are proposed to improve the performance of traditional RL methods such as Q-learning. The proposed improvements are: the addition of a forgetting mechanism, the use of feature-based state inputs, and the addition of a hierarchical structure to an RL agent. Experimental results are presented and utilized in evaluation of the proposed methods. The proposed algorithms are compared to established methods and to theoretically optimal solutions.