Decentralized Scalar Field Mapping using Gaussian Process

Abstract

Decentralized Gaussian process (GP) methods provide a scalable framework for multi-agent scalar-field estimation by allowing each agent to maintain a local GP over its assigned subdomain. Because neighboring agents condition their GP models on different measurements, their GP posteriors generally disagree over overlapping regions. This paper introduces the inducing points decentralized Gaussian process (IPD-GP) framework, which exploits these discrepancies to improve posterior consistency over shared regions. Each sender constructs receiver-specific data packets by selecting inducing locations targeted at the pairwise overlap and transmitting the corresponding posterior means and variances. Each receiver then selectively assimilates informative packets as persistent fictitious measurements in its local GP. The developed architecture preserves decentralized computation and communication without requiring full dataset exchange or centralized inference. Performance is evaluated using normalized root-mean-square error, negative log predictive density, and their overlap-restricted counterparts, with comparisons against non-communicating local GPs, decentralized aggregation methods, and a centralized GP. Simulation results show that IPD-GP achieves competitive predictive accuracy while reducing inter-agent communication.