Bayesian optimization algorithms
Dr. Bedřich Sousedík, Department of Mathematics and Statistics.
Our project focuses on the development and evaluation of Bayesian optimization algorithms for high-dimensional black-box optimization problems. In particular, we are designing new surrogate modeling and acquisition strategies intended to improve optimization efficiency and scalability in challenging, high-dimensional settings.
A major component of the project involves benchmarking our solver against existing state-of-the-art Bayesian optimization methods using several established benchmark suites. These include: Lasso-DNA (180 dimensions), MOPTA08 (124 dimensions), Humanoid (6392 dimensions), as well as additional synthetic and real-world optimization tasks.
The experiments require repeated large-scale evaluations across multiple random seeds and competing methods, involving substantial computational workloads for surrogate model training, acquisition optimization, and parallel objective evaluations. Some of the benchmarks, particularly the Humanoid task(bare minimum recommended is 12-core for this in particular), also involve computationally intensive simulation-based evaluations.