New Holonic Active Distillation Framework Enables Scalable Multi-Agent Learning in Multi-Sensor Systems

A team led by Dani Manjah and Tim Bary introduces Holonic Active Distillation (HAD) to tackle scalability and adaptability in multi-sensor networks. The approach operates within a Holonic Multi-Agent System (HMAS), which organizes agents hierarchically into holons. Their Clustered Stream-Based Active Distillation (CSBAD) framework deploys specialized student models that gather local sensor data, query pseudo-labels from teacher models, and cluster similar sensors together. This enables the system to dynamically handle sensors joining or leaving without retraining from scratch, balancing local specialization with global generalization.

The paper, accepted at EMAS 2025, demonstrates that the holonic organization effectively adapts to sensor departures and re-integrations while maintaining performance. The authors also analyze trade-offs between incremental model updates, system reorganization, and scalability limits. Key challenges identified include model drift and long-term adaptation, but the framework shows promise for open, evolving sensor networks. This research has practical implications for IoT, smart cities, and industrial monitoring where sensor configurations change frequently.