Gaussian processes for global optimization

Conference item

We introduce a novel Bayesian approach to global optimization using Gaussian processes. We frame the optimization of both noisy and noiseless functions as sequential decision problems, and introduce myopic and non-myopic solutions to them. Here our solutions can be tailored to exactly the degree of confidence we require of them. The use of Gaussian processes allows us to benefit from the incorporation of prior knowledge about our objective function, and also from any derivative observations. Using this latter fact, we introduce an innovative method to combat conditioning problems. Our algorithm demonstrates a significant improvement over its competitors in overall performance across a wide range of canonical test problems.

Authors: Osborne, MA; Garnett, R; Roberts, SJ

• Publication status: Accepted
• Peer review status: Peer reviewed
• Publication date: 2009-01-18
• Event title: Third International Conference on Learning and Intelligent Optimization (LION3)
• Event location: Trento, Italy
• Event website: http://www.wikicfp.com/cfp/servlet/event.showcfp?eventid=2951&copyownerid=1896
• Event start date: 2009-01-14
• Event end date: 2009-01-18
• Language: English
• Keywords: decision theory; Gaussian processes; Bayesian methods; global optimization; FFR; noisy optimization; non-convex optimization