An improved Q-learning algorithm for an autonomous mobile robot navigation problem

This work applies the popular reinforcement learning methodology of Q learning in a typical robot control navigation problem. It is a two dimensional (2D) set-up where a robot tries to learn its path through its environment by avoiding any obstacles that may be encountered on its way from its home to a final destination (a goal state). During the navigation, trajectory of all the state-action pairs is stored and is replayed in a backward direction to propagate the refined Q values from any state to a goal state. This effort greatly reduces the convergence rate for the Q-table as the results obtained from the simulations indicate an excellent level of performance once compared with the traditional Q-learning.