APF-DQN: Adaptive Objective Pathfinding via Improved Deep Reinforcement Learning Among Building Fire Hazard

Ke Zhang, Dandan Zhu, Qiuhan Xu, Hao Zhou, Xuemei Peng

Abstract: Evacuation path planning is a critical task to enable the safety of individuals in a fire hazard. Current evacuation planning approaches mainly calculate a fixed optimal path given a deterministic task. Nevertheless, fire evacuation guidance confronts some vital challenges including multi-exits existing in building and unstable evacuation path caused by dynamic fire spread. To resolve these issues, this paper proposes an evacuation agent which possesses a novel Artificial Potential Field Deep Q-Learning (APF-DQN) algorithm to calculate an evacuation route which evacuation agent enables choosing an appropriate exit and plan a dynamic evacuation path. A concept called artificial potential field is introduced into deep Q-learning architecture to lead agent adaptively choose targeted exit and avoid damage from fire spread. Meanwhile, deep Q-learning framework ensures evacuation agent plan a dynamic path. Then, APF-DQN is estimated in a proposed simulation experiments and compared with several conventional path-finding methods. Our APF-DQN reduces time-step cost by 18.7% and increases distance to closest fire by 20.1% compared with classical A star and APF methods. Our code can be downloaded from URL: https://github.com/ColaZhang22/APFDQN-Indoor-fire-hazard-path-planning.