Developer builds explainable AI system to optimize wildfire evacuation decisions

A developer exploring crisis AI created an Explainable Causal Reinforcement Learning (XC-RL) framework designed to improve decision-making in wildfire evacuation logistics. The project was inspired by a personal experiment in which a trained RL agent repeatedly routed virtual evacuees toward a fire due to a flawed reward function. Unlike standard tools such as SHAP or LIME, which identify influential inputs but not causal relationships, the new system uses a Structural Causal Model to encode true cause-and-effect dynamics like fire spread and road closures. The architecture combines a causal world model, a PPO-based policy network, and an inverse simulation verifier that runs counterfactual scenarios to confirm whether the AI's stated reasons for actions match its actual behavior. The framework aims to address the opacity of traditional reinforcement learning in safety-critical situations where understanding why a decision was made is as important as the decision itself.