Adaptive sampling for learning gaussian processes using mobile sensor networks.

Area of Science:

• Robotics and Artificial Intelligence
• Statistical Modeling
• Geophysics and Environmental Science

Background:

• Gaussian processes are effective for modeling spatio-temporal data but require accurate covariance functions.
• Estimating unknown covariance functions from noisy measurements is a significant challenge in data assimilation.
• Self-organizing sensing agents offer a potential solution for adaptive data acquisition.

Purpose of the Study:

• To develop self-organizing sensing agents capable of adaptively learning anisotropic, spatio-temporal Gaussian processes.
• To improve the estimation of the covariance function using noisy measurements and agent movement.
• To introduce an optimal sampling strategy for efficient data acquisition.

Main Methods:

• Utilizing a novel class of anisotropic covariance functions for Gaussian processes.
• Employing a Maximum A Posteriori (MAP) probability estimator for covariance function estimation.
• Implementing an optimal sampling strategy based on minimizing the Fisher Information Matrix.

Main Results:

• Demonstrated the effectiveness of self-organizing agents in learning complex spatio-temporal patterns.
• Showcased the adaptability of the agents in improving covariance function estimation with noisy data.
• Validated the proposed sampling strategy for efficient information gathering.

Conclusions:

• The proposed approach enables adaptive learning of spatio-temporal Gaussian processes by self-organizing agents.
• Agent movement significantly enhances the quality of the estimated covariance function.
• This method provides a robust framework for intelligent data acquisition in physical phenomena modeling.