Tuesday 15 December 2020

Thought Leadership Event: Dr Sandy Brownlee on "Improving trust in the results of search-based optimisation"

 Date: 21st January 2021, 14:00 - 15:00

Speaker: Dr Sandy Brownlee, University of Stirling
TitleImproving trust in the results of search-based optimisation 

The event was hosted by BT Applied Research's Dr Gilbert Owusu and his team.

Search-based methods, including techniques like metaheuristics, local search, and evolutionary algorithms, are powerful tools for finding good solutions to difficult optimisation problems. Usually the problem is formulated as a representation (what a solution looks like) and a fitness function (which measures the quality of solutions). The algorithm will explore the space, generating new solutions through a variety of randomised processes, and eventually settle on a good solution. Sometimes the solutions found by these processes can be a little unexpected or "outside the box", which can make acceptance difficult. 

In this talk, I looked at a few approaches I've taken to using visualisation for trying to improve trust in the solutions of search-based optimisation. These are about trying to understand what the algorithm has learned about the problem during the search, and how robust the solutions it found really are. 

The example applications are a software tuning problem and some problems in building design, but the general principles should be much more widely applicable. 

Dr Sandy Brownlee joined the Division of Computing Science and Mathematics at the University of Stirling in 2013, where he is currently a Lecturer. He gained BSc and PhD degrees from Robert Gordon University in 2005 and 2009 respectively, and has also worked at Loughborough University, and as a software engineer in the energy industry. He is interested in "explainable" optimisation: techniques that find good solutions for optimisation problems but also reveal underlying information about the problem to help people make informed decisions. 

His main focus is in metaheuristics, including evolutionary algorithms and estimation of distribution algorithms; related issues such as fitness modelling (and mining such models), handling constraints and multiple objectives, decision support, and what makes particular algorithms suited to particular problems. This work has resulted in over 60 publications in peer-reviewed venues, and has found applications in areas including scheduling and simulation-based optimisation in civil engineering and transport, software engineering, healthcare, and art. 

No comments:

Post a Comment

Note: only a member of this blog may post a comment.