Waseem Gharbieh and Amjed Al-Mousa
Neural networks, obstacle avoidance, partially observable, linearregression, dynamic environment, reinforcement learning
In this paper, a decision-making system that is based on feature ranking is presented to allow a robot to navigate back and forth between two points. The path of the robot contains an unknown number of both static and moving obstacles that get in its way. To simulate the use of real sensors, the implementation uses only local information available within a predefined radius for decision- making. To address this problem, feature ranking is introduced, where the top k obstacles are taken into consideration in planning the robot path. Feature ranking is applied to linear regression and neural networks to compare their generalization ability against a baseline policy. During the development of the system, the robot undergoes a learning phase, where the moving obstacles move at 50% of the robot’s speed. When learning phase is complete, the generalization ability of the policy learned by both algorithms is tested by generating environments with various obstacle speeds. The results show that the policy developed performs much better than the baseline policy with the neural network outperforming linear regression in the majority of the environments.
Important Links:
Go Back
