Reinforcement learning path planning algorithm based on obstacle area expansion strategy

Abstract

We improve the traditional Q($$ \lambda $$)-learning algorithm by adding the obstacle area expansion strategy. The new algorithm is named OAE-Q($$ \lambda $$)-learning and applied to the path planning in the complex environment. The contributions of OAE-Q($$ \lambda $$)-learning are as follows: (1) It expands the concave obstacle area in the environment to avoid repeated invalid actions when the agent falls into the obstacle area. (2) It removes the extended obstacle area, which reduces the learning state space and accelerates the convergence speed of the algorithm. Extensive experimental results validate the effectiveness and feasibility of OAE-Q($$ \lambda $$)-learning on the path planning in complex environments.