HiDRA-Evac: Hierarchical Dynamic Cluster Reassignment With Profile-Aware Agents for Crowd Evacuation Using Multi-Agent Reinforcement Learning
作者
M Karthik Krishna,K. Naga Srinivasan,A Hariharan,V Pranav,S Chitrakala
出处
期刊:日期:2026-03-26卷期号:: 1-6
标识
DOI:10.1109/icdsaai69492.2026.11505094
摘要
Multi-Agent Reinforcement Learning (MARL) has emerged as a promising paradigm for learning coordinated evacuation policies directly from interaction. Recent evacuation frameworks based on GRU-enhanced Multi-Agent Proximal Policy Optimisation (GRU-MAPPO) demonstrate strong coordination through hierarchical leader–follower structures. However, most existing approaches rely on static crowd grouping, which is poorly suited to dense, dynamically evolving scenarios, and make unrealistic assumptions about homogeneous agent behaviour. HiDRA-Evac investigates adaptive crowd grouping, dynamic cluster reassignment (DCR), and agent personalisation within a GRU-MAPPO-based evacuation framework. K-Means and DBSCAN-based clustering strategies are compared, with and without DCR. Experimental results show that DBSCAN combined with DCR consistently yields superior evacuation efficiency, reduced congestion, and improved learning stability compared to static grouping, highlighting the importance of adaptive grouping and heterogeneous agent modelling for realistic MARL-based crowd evacuation.