Helping Domain Experts Build Better Algorithms
28"Helping Domain Experts Build Better Algorithms: Automated Performance Modelling, Configuration, and Selection"
Google Tech Talk
August 8, 2011
Presented by Frank Hutter.
ABSTRACT
Algorithm developers and end users in a wide variety of areas often face questions like the following:
Which parameter setting should I use to optimize my algorithm's empirical performance?
Which algorithm components are most critical to achieve good performance?
Which of two (or more) available algorithms will perform best on a given new instance?
We describe fully formalized domain-independent methods that aim to answer these problems based on machine learning and optimization techniques. We illustrate the power of these automated methods by optimizing state-of-the-art solvers for two fundamental problems: propositional satisfiability (SAT) and mixed integer programming (MIP). With minimal human effort, in several cases our methods sped up the best existing SAT and MIP solvers by orders of magnitude.
Based on joint work with Holger Hoos and Kevin Leyton-Brown.
Speaker Info:
Frank Hutter http://www.cs.ubc.ca/~hutter/