PATH PLANNING METHOD OF DEEP REINFORCEMENT LEARNING WITH IMPROVED REWARD FUNCTION

Aimed at the sparse reward problem of deep reinforcement learning in path planning, a deep reinforcement learning model based on potential reward function is proposed. By designing a new reward function, the model improved the reward density and sample utilization, reduced the difficulty of training, and improved the success rate of agent in different maps. The simulation results show that the path planning success rate of the improved model is improved by 7.08 percentage points on simple maps and 12.60 percentage points on complex maps. Compared with the most advanced algorithms, the routing success rate is similar, but the length of the planned path result is shortened.