Daniël Fokkinga joined the ADA research group as a Master’s student in September of 2018. His research is jointly supervised by Anna Louise Latour and Marie Anastacio and applies Automated Algorithm Configuration (AAC) on a pipeline for solving Stochastic Constraint Optimisation Problems (SCOPs).
An example of such a SCOP is the viral marketing problem. We are given a social network where an edge between two nodes represents a probabilistic communication relationship between two people in the network. We want to launch a new product by handing out a limited number of free samples to people in the network. We hope that they will like our product and spread the word to turn their acquaintances into buyers of our product. To whom in the network do we give the free samples, in order to maximise the expected number of people that will buy our product after they’ve heard about it from others? Who are the most influential people in our network?
Recently, a new pipeline for solving SCOPs was proposed. While it has shown its merit in a proof of principle, the different steps in the pipeline have not been optimised yet. While alternative design choices for elements in the pipeline have been proposed, their influence on the performance of the pipeline over a wide range of problems has not been extensively explored.
Daniël’s focus is on exposing these alternatives as parameters to optimise the pipeline for different applications, using Automated Algorithm Configuration.
Automated Algorithm Configuration allows a user to leave the tuning of parameters and thus the decision of the best approach to follow for parts of the full pipeline to a computer. Given a set of example instances, a configurator finds the combination of parameter values that is most likely to perform well on a new instance.
With this research Daniël hopes to show another successful application of Algorithm Configuration in a new field and improve the performance of previously developed algorithms.