Rikesh Yadav
Back to Projects
Manuscript in PreparationCurrent Research

Adaptive Evacuation: A GNN-Enhanced Deep Reinforcement Learning Framework

Deep Reinforcement LearningGraph Neural NetworksIoTPyTorch Geometric

Developing a hybrid Graph Neural Network and Deep Reinforcement Learning framework for adaptive indoor emergency response. This approach bridges theoretical RL with real-time execution in IoT-enabled environments.

Research Motivation

Current evacuation systems rely on static routing strategies that fail to account for the dynamic, rapidly changing nature of emergency conditions (e.g., spreading fire, collapsing infrastructure). A dynamic, self-learning system is required.

Problem Statement

How can autonomous systems calculate and broadcast optimal evacuation paths in real-time when the underlying spatial network is actively degrading?

Methodology & Architecture

We map the indoor environment to a graph structure where nodes represent physical locations and edges represent traversable paths. Edge weights dynamically update based on IoT sensor data. A Graph Convolutional Network (GCN) extracts spatial features, which are then passed to a Deep Q-Network (DQN) to learn optimal evacuation policies.

System Architecture

The architecture consists of three layers: the Physical Layer (IoT sensors), the Representation Layer (GCN embeddings), and the Decision Layer (DQN agent).

Experiments

Simulated emergency scenarios in Gymnasium, comparing the GCN-DQN agent against standard Dijkstra and basic tabular Q-learning across maps of varying complexity.

Results & Findings

The hybrid model demonstrated a 40% reduction in average evacuation time and a 99% success rate in high-volatility environments compared to static baseline routing.

Discussion

Future Work

Extending the framework to Multi-Agent Reinforcement Learning (MARL) to account for crowd dynamics and pedestrian bottlenecks.