Multivariate Active Learning and Adaptive Sampling with Multi-Kernel Gaussian Processes
Thien Hoang Nguyen, Nathan Wallace, Nicholas Harrison, and Salah Sukkarieh
AI summary
Problem
Traditional environmental sampling is manual and costly, while existing robotic methods typically focus on single variables or make suboptimal assumptions that fail to capture complex inter-variable correlations or optimize travel efficiently.
Approach
The authors introduce a Multi-kernel Gaussian Process (MKGP) to model spatial distributions and cross-correlations across multiple quantities of interest, combined with a Multivariate Adaptive Sampling (MVAS) algorithm that selects optimal sampling locations by balancing information gain against travel cost.
Key results
- MKGP accurately captures spatial distributions and complex correlations across multiple environmental variables
- MVAS efficiently plans sampling routes that maximize information gain while minimizing travel cost
- System achieves lower root-mean-square error than single-output and single-kernel Gaussian process baselines
- Explicitly leveraging correlated prior data significantly improves target variable mapping accuracy
Why it matters
Enables precision agriculture and environmental monitoring to reduce chemical use and operational costs through faster, more accurate robotic mapping.
Abstract
In agriculture, understanding the distribution and relationship between different aspects of the environment is im- portant for minimizing chemical use and reducing environmental impact. Traditionally, it is done by manually collecting samples on the field and then sending them to a laboratory for analysis. This is not only labor-intensive and costly, but the results will still be outdated. There is thus a growing interest in developing robotic systems to map these variables and uncover their correlations in real time. However, existing learning and sampling methods only focus on one quantity of interest (QoI) or make assumptions that might lead to sub-optimal results when there are multiple QoIs. In this work, we propose a multivariate active transfer learning and intelligent adaptive sampling system that can simultaneously learn the most accurate models for multiple QoIs as well as the relationship between them, and leverage that knowledge to select the next best locations to sample. Performance benchmarking against existing methods shows that QoIs are mapped more accurately, complex correlations between QoIs are identified more precisely, and travel routes are planned more efficiently.