Comparison of deep reinforcement learning algorithms: Path Search in Grid World

Deep reinforcement learning is being used to teach robots elaborate and complex tasks, and recently, it has gained successful results in various fields of robotics. However, there is still a limit to directly applying the existing deep reinforcement learning algorithms due to the continuous and high-dimensional states and actions inherent in the robot, and a method of inducing correct learning through proper expression of the environment in which the robot will work is required. As a way to enable the robot to recognize the work environment, there is observation through a camera, and a mapping from this to the grid world is a way to express the work environment. This allows us to considerably reduce the dimensions of state and action and find the optimal policy for the path search problem using deep reinforcement learning algorithms. In this paper, we compare the simulation results of various deep reinforcement learning algorithms for path search, which is a representative problem depicted as a grid world, and present the environment, model architecture and parameters used in the simulation.