Many modern science and engineering applications, such as machine learning, hyperparameter optimization of neural networks, robotics, cyber-physical systems, etc., call for modeling techniques to model black-box functions. Gaussian Process (GP) modeling is a popular Bayesian non-parametric framework heavily employed to model expensive black-box functions for analysis such as prediction or optimization.
Traditionally, GP models assume stationarity of the underlying, unknown function. As a result, a unique covariance kernel (with constant hyperparameters) can be used over the entire domain. However, many real-world systems such as cyber-physical systems and controls often require modeling and optimization of functions that are locally stationary and globally non-stationary across the domain. Thus, with many current GP modeling frameworks operating under assumptions of stationarity of an underlying, unknown function, there is a need for a modeling technique that makes no such assumptions.
Researchers at Arizona State University have developed a modeling technique for black-box functions. In particular, a Gaussian Process (GP) modeling technique for unknown, globally non-stationary heterogeneous functions that can be applied in various science and engineering applications.
Related publication: Class GP: Gaussian Process Modeling for Heterogeneous Functions
Potential Applications:
- Black-box modeling for science and engineering applications:
- Machine learning
- Neural networks
- Robotics
- Cyber-physical systems
Benefits and Advantages:
- Modeling and optimization of unknown, globally non-stationary heterogeneous functions
- Applicable for black-box modeling for many modern-day science and engineering applications