Performance evaluation of double action Q-learning in moving obstacle avoidance problem

This paper describes the performance evaluation of double action Q-learning in solving the moving obstacle avoidance problem. The evaluation is focused on two aspects: 1) obstacle avoidance; and 2) goal seeking; where four parameters are considered, namely, sum of rewards, no. of collisions, steps per episode, and obstacle density. Comparison is made between the new method and the traditional Q-learning method. Preliminary results show that the new method has the sum of rewards (negative) 29.4% and 93.6% less than that of the traditional method in an environment of 10 obstacles and 50 obstacles respectively. The mean no. of steps used in one episode is up to 26.0% lower than that of the traditional method. The new method also fares better as the number of obstacles increases.