Human-Aligned Decision Transformers for wildfire evacuation logistics networks for low-power autonomous deployments

Short summary

Decision Transformers combined with human-aligned reward functions enable safer evacuation route planning by conditioning on historical trajectories and expert-annotated safety rules. The approach adds guardrails to prevent dangerous shortcuts while optimizing speed and safety in non-stationary fire environments. Implementation uses context-aware attention masking and multi-objective RTG to handle real-world constraints.

• •Applies Decision Transformers to wildfire evacuation routing with human safety alignment via learned penalty networks
• •Uses offline learning with historical evacuation data and simulated scenarios to avoid real-world training risk
• •Implements context-aware attention masking and human-aligned RTG to prioritize safety over pure computational optimization

Generated with AI, which can make mistakes.