Emergency fire escape path planning model based on improved DDPG algorithm

Currently, there is a growing demand for fire emergency evacuation capability of buildings. However, current plans based on pre-established static evacuation routes, fail to consider the dynamic nature of real-time fire scenarios. The objective of this paper is to propose a path planning model for fire emergency scenarios. The model is designed to provide evacuation paths for stranded individuals based on fire-related factors. Firstly, the model acquires relevant data through fire simulation, defines the reward function, and establishes an environment for training the agent. Secondly, the Deep Deterministic Policy Gradient (DDPG) algorithm is improved for the characteristics of the fire scenario. The intrinsic motivation was introduced to the DDPG to enhance its ability to explore the state space, and the hindsight experience replay strategy was implemented to improve the algorithm's training effectiveness. For the hyperparameter sensitivity problem in reinforcement learning, the Beluga Whale Optimization algorithm was employed for hyperparameter optimization. The experimental results show that the reinforcement learning model can plan evacuation paths starting from any location, and the final paths effectively balance the trade-off between risk and distance cost in the environment. The improved DDPG algorithm converges to increase the average reward by about 100, and the applicability of the model is verified in different fire environments. The application of the model can provide a good basis for the selection of emergency evacuation paths, and it has theoretical and practical value for the designation of scientific evacuation plans and the evaluation of the performance of building emergency systems. •Established a fire evacuation mission training environment for agent.•Improved exploration and empirical replay strategies in the DDPG algorithm.•Hyperparameters were optimized using swarm intelligence algorithms.•The effectiveness of the method has been validated in a variety of tasks.