Integrating dynamic environment modelling with reinforcement learning for fire safety design

Escape path planning belongs to the domain of fire safety design in buildings, ensuring the safety of individuals during emergencies. Applied Reinforcement Learning (RL) research has focused on optimizing training for simple evacuation scenarios. Research findings indicate that agents can learn to escape environments using path-planning methods. However, it has yet to be considered in depth how immediate changes to the environment affect the learning process. By using dynamic environment modeling, the agent can learn to modify floor plans based on the results of escape paths. This approach leads towards a performance-based validation tool, where the results of evacuation path planning can be used to, e.g., support regulatory compliance checks. Our work focuses on creating a reinforcement learning environment that simulates a floor plan to train the agent to identify escape paths and strategically optimize the layout by placing the exits for evacuation. The environment creation is based on a grid-based environment using a discrete action space. This allows for fast prototyping since learning in environments with fewer choices leads to less variability of outcomes and facilitates comparison with existing research, which the study addresses regarding reproducibility and comparability of the environments and the robustness of the algorithm's outcomes.