Q-Learning Framework as a Solution for an Obstacle Avoidance Problem in Unknown Environment

Abstract

This paper presents machine learning approach as a solution for an obstacle avoidance problem. Q-learning, as one of the reinforcement learning algorithms, imparts the learning process based on trial and error and a corresponding reward into the behaviour of an intelligent agent - a mobile robot. The adaptable actions of a mobile robot in situations when that behaviour is necessary are the main advantage over conventional methods for designing a navigational path. The implemented algorithm characterizes simplicity and efficiency, and certainty in terms of reaching optimal behaviour after the certain number of learning episodes. Experimental results show proper exploration strategy with gradually improving mobile robot state to action mapping by adjusting Q-value function in a described manner. With more episodes conducted this adaptable control system could lead to a fully autonomous mobile robot, which is one of the main demands in modern intelligent manufacturing systems in which stochastic changes in the environment can results with failure in the entire production process.