Keynote Lecturer: Michele Sebag
In many domains such as machine learning, constraint satisfaction or stochastic optimization, algorithm portfolios have been designed to handle the diversity of problem instances. In order to get peak performance on any particular problem instance, one must be able to select the algorithm and hyper-parameter setting best suited to this problem instance. The efficient deployment of algorithmic platforms outside research labs thus depends on the appropriate selection of the algorithm and hyper-parameter setting.
The talk will show how the issue of automatic algorithm selection and configuration can be handled by exploiting the portfolio archive, recording which algorithms and settings have been used on problem instances, and the corresponding result. Taking inspiration from the Matchbox system proposed by Stern et al. (2010), this issue is tackled as a collaborative filtering problem: each problem instance gives "marks" to some algorithms, and algorithms with better performance on this problem instance get better marks. Collaborative filtering can thus be exploited to recommend algorithms for a given problem instance. The talk will focus on the cold-start problem (how to deal with a new problem instance), presenting the algorithm recommender system ALORS, with applications in SAT, gradient-free optimization and machine learning.
Some perspectives about the exploitation of the ALORS system in order to propose a typology of problem instances and algorithms will be discussed.
Joint work with Mustafa Misir, Rémi Bardenet, Balazs Kégl