DEMO Se­mi­nar: two talks on mul­tiob­jecti­ve op­ti­mization


25.4.2019 9:00 — 12:00

Sijainti: Mattilanniemi, Agora Ag C421.1
DEMO - Decision Analytics utilizing Causal Models and Multiobjective Optimization is a profiling area of our university, This presentation is aimed at anyone interested in themes of DEMO and does not require deep knowledge of e.g. causality or statistics.

Speaker: Dr. Tinkle Chugh (Uni­ver­si­ty of Exe­ter, UK) 

Tit­le: A mul­tiob­jecti­ve op­ti­mization ap­proach in buil­ding Gaus­sian process mo­dels

Gaussian processes (GPs) have been widely used in optimization and machine learning communities. Some of the problems where GPs have gained their popularity are non-linear regression, classification and Bayesian optimization (both single and multiobjective). The main advantage of GPs is that they provide a predictive distribution of the data instead of a point prediction. The uncertainty provided by the distribution can further be used in decision making and optimizing an acquisition function in Bayesian optimization. Despite their wide applicability, a little attention has been paid to the problem of selecting hyperparameter values and kernel functions. Several options exist and it is not straightforward to select a particular value of hyperparameter and a kernel function. In this talk, I will present a multiobjective optimization approach in building Gaussian process models which addresses the challenges of selection of hyperparameter values and different kernels.

Title: A stu­dy on using dif­fe­rent sca­la­rizing func­tions in Baye­sian mul­tiob­jecti­ve op­ti­mization

Scalarizing functions have been used for many decades in Multiple Criteria Decision Making (MCDM) community for converting a multiobjective optimization problem into a single objective optimization problem. In the last few years, their use in evolutionary multiobjective optimization (EMO) has also been increased especially in decomposition based algorithms. However, their use in solving expensive multiobjective optimization is scarce and only few studies exist. In this talk, I will present a review of different scalarizing functions from both MCDM and EMO communities and their use in Bayesian multiobjective optimization when solving expensive multiobjective optimization problems. The results on different benchmark problems clearly show a correlation between the performance of different functions and the fitness landscape created by them.

You are warmly welcome!

More information
Kaisa Miettinen
Professor, director of DEMO
Faculty of Information Technology
050 373 2247