Distributed Calibration of Agent-based Models
Keywords: Agent-based Modeling, Distributed Machine Learning, Decentralized Data, Epidemiological Modeling
TL;DR: We introduce a distributed learning protocol that enables calibrating agent-based models with sensitive data from multiple institutions, without sharing the raw data, by splitting the process between clients and a server.
Abstract: Agent-based models (ABMs) simulate complex systems by modeling the interactions between individual agents. Calibrating ABMs to real-world data is critical for their practical utility, but is hindered by the fact that granular data is often siloed across institutions due to privacy concerns. We propose a new protocol for distributed calibration of ABMs that allows institutions to collaborate on model calibration without sharing raw data. The protocol splits the calibration neural network (CalibNN) between the data clients and a central server. Each client generates embeddings from their local data and transmits them to the server, which merges the embeddings to calibrate the ABM. Gradients are propagated back to the clients to update their local models. On preliminary experiments simulating the COVID-19 pandemic, we find the distributed protocol achieves calibration accuracy on par with centralized calibration using pooled data. This demonstrates the potential to leverage sensitive data to improve ABMs while preserving privacy.
Submission Number: 7
Loading