B0942
Title: A novel improvement acquisition function
Authors: Raju Chowdhury - Indian Institute of Technology Madras (India) [presenting]
Neelesh Shankar Upadhye - Indian Institute of Technology Madras (India)
Abstract: In the domain of Bayesian optimization (BO), one of the most common acquisition functions is the expected improvement (EI). Despite abundant resources for solving unconstrained problems, EI lacks the structures necessary to address constrained problems. An improvement-based acquisition function is proposed for solving constrained optimization problems. Like all other acquisition functions, the proposed one balances the trade-off between exploring and exploiting the search space. It also keeps a balance between the feasible and infeasible regions, this is important given that constrained problems are dealt with. Regardless of starting from an infeasible or feasible point, the method finds a feasible optimal solution sooner than existing methods. On benchmark problems, the method is compared to existing methods.
