Gaussian Processes for Multi-agent Systems

Gaussian Processes (GPs) are flexible non-parameteric bayesian models which allow us to learn an unknown phenomenon in a wide variety of fundamental signal processing problems, for which a data model may not be readily be available. In partticular, for multi-agent systems, these include, for example, field estimation using multi-agent systems, trajectory tracking, nonlinear model predictive control or sensor localization. Despite the broad applicability, GPs suffer from significant computational burden particularly for larger datasets with a computational complexity of O(N³). There are numerous ways to overcome this impediment e.g., using low rank approximation methods, distributed optimization or exploiting domain-awareness. This MS thesis will explore and advance one of these directions, for multi-agent systems. See references below.  

[1] Zhai, P., & Rajan, R. T. (2023, June). Distributed Gaussian Process Hyperparameter Optimization for Multi-Agent Systems. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) (pp. 1-5). IEEE.

[2] Mishra, A., & Rajan, R. T. (2024). Domain-Aware Gaussian Process State-Space Models. Available at SSRN 4955025.

[3] Ban, H., Riemens, E. H., & Rajan, R. T. (2024, August). Malleable Kernel Interpolation for Scalable Structured Gaussian Process. In 2024 32nd European Signal Processing Conference (EUSIPCO) (pp. 997-1001). IEEE.

[4] Williams, C., & Rasmussen, C. (1995). Gaussian processes for regression. Advances in neural information processing systems, 8.

Requirements

A self-motivated student with a strong background system design, linear algebra and statistical signal processing. Previous experience distributed algorithms and project related to navigation is an added plus. Advanced coding skills in MATLAB and Python (or R) is required. Strong written and verbal communication skills in English is mandatory. The expected project duration is 8-12 months.

Contact

Last modified: 2025-02-12