Abstract
A directed graph based evolutionary algorithm called Genetic Network Programming (GNP) has been proposed and extended with reinforcement learning algorithm (GNP-RL). Previously, GNP and GNP-RL had been applied successfully into the generation of programs for controlling mobile robot behavior and good training results were obtained. This paper aims to analyze the robustness of GNP-RL in the testing phase. For this purpose, programs for wall-following behavior of Khepera robot are generated by using both GNP and GNP-RL. In order to measure the robustness, variations in the operating environment is needed. In this paper, faulty sensors in the testing period emulate the change of the operating environment. The robustness of the evolved program is important to cope with the new situation. By using the reinforcement learning algorithm and ε-greedy policy in the testing period, a relatively good robustness against environment change is achieved by GNP-RL with proper parameter settings. After the environment changes, GNP-RL is still able to obtain rewards although the sensor(s) supplies wrong input values. GNP-RL has the flexibility to automatically and easily adapt to operating environment changes. It is done by selecting alternative actions which are prepared during the training period. In the simulation part, Khepera robot is simulated on Webots. The results obtained from the training and two different testing mechanisms are presented and analyzed. Additionally, the effects of different parameter settings of GNP- RL in the training and testing phase are analyzed to find the optimum settings in a given environment.
