Reinforcement learning for pathfinding with restricted observation space in variable complexity environments

Abstract

View Video Presentation: https://doi.org/10.2514/6.2021-1755.vid Path finding is a common problem in computer science that has applications in robotics and autonomous systems and can be solved using reinforcement learning. Many solutions to path finding problems rely on the agent having full knowledge of its environment, however this may limit the agent from being able to act effectively in environments it has not been trained on. We modified Open AI’s Frozen Lake, a simple 2D path finding environment, to include a limited observation space. Using a variation of Q – learning, Modified Connectionist Q-Learning, an agent was trained on Frozen Lake and our new environment and tested on a series of maps it had not been trained on, as well as on a dynamic map. The agent trained on the modified,limited observation space environment was able to generalize its training to solve maps it had not seen before, in contrast to the agent trained on the Frozen Lake environment, which was not able to generalize its training.