21: Dynamic Surrogate Modeling for Online Multiobjective Optimization

Music City Center 

Finding decisions in real-time that optimize multiple outcomes involves sequential decision-making with a priori uncertainty in multivariate functional objectives. In addition, in practice observations of functional objectives can also be difficult to obtain due to computational or monetary expense. This challenge is common in engineering applications, such as optimizing a racing vehicle's performance by minimizing both lap time and fuel consumption. To address this, we propose a dynamic surrogate modeling strategy based on Gaussian Process (GP) regression with auto-regressive (AR) structures to both emulate and forecast objective functions. This strategy both captures temporal dependencies while efficiently approximating time-varying objective functions. The AR-GP model is incorporated into a multi-objective online optimization framework, allowing for adaptive decision making with reduced computational overhead. Preliminary results, based on an example motivated by engineering design and autonomous vehicle applications, demonstrate that AR-GP modeling outperforms standard GP surrogates by providing more accurate predictions and better uncertainty quantification.

Gaussian Process

Dynamic Models

Online Optimization

Surrogate Modeling

Multiobjective Optimization

Emulation 

Section on Physical and Engineering Sciences