RoamML distributed continual learning: Adaptive and flexible data-driven response for disaster recovery operations

Abstract

In the aftermath of natural disasters, Human Assistance & Disaster Recovery (HADR) operations have to deal with disrupted communication networks and constrained resources. Such harsh conditions make high-communication-overhead ML approaches — either centralized or distributed — impractical, thus hindering the adoption of AI solutions to implement a critical function for HADR operations: building accurate and up-to-date situational awareness. To address this issue we developed Roaming Machine Learning (RoamML), a novel Distributed Continual Learning Framework designed for HADR operations and based on the premise that moving an ML model is more efficient and robust than either large dataset transfers or frequent model parameter updates. RoamML deploys a mobile AI agent that incrementally train models across network nodes containing yet unprocessed data; at each stop, the agent initiate a local training phase to update its internal ML model parameters. To prioritize the processing of strategically valuable data, RoamML Agents follow a navigation system based upon the concept of Data Gravity, leveraging Multi-Criteria Decision Making techniques to simultaneously consider many objectives for Agent routing optimization, including model learning efficiency and network resource utilization, while seamlessly blending subjective insights from expert judgments with objective metrics derived from quantifiable data to determine each next hop. We conducted extensive experiments to evaluate RoamML, demonstrating the framework’s efficiency to train ML models under highly dynamic, resource-constrained environments. RoamML achieves similar performance to centralized ML training under ideal network conditions and outperforms it in a more realistic scenario with reduced network resources, ultimately saving up to 75% in bandwidth utilization across all experiments.