A Preliminary Evaluation of Machine Learning in Algorithm Selection for Search Problems
Lars Kotthoff, Ian P. Gent, Ian Miguel
larsko@cs.st-andrews.ac.uk
University of St Andrews
The Algorithm Selection Problem
Methodology
Methodology
Methodology
Methodology
-
5 different kinds of Machine Learning
- case-based reasoning
- classification
- regression on runtime
- regression on log of runtime
- statistical relational learning
Methodology
-
no parameter tuning
- Machine Learning algorithms
- search algorithms
Methodology
-
no parameter tuning
- Machine Learning algorithms
- search algorithms
- no subset selection of search algorithm portfolio
Methodology
-
no parameter tuning
- Machine Learning algorithms
- search algorithms
- no subset selection of search algorithm portfolio
- no evaluation of different search problem encodings
Methodology
-
no parameter tuning
- Machine Learning algorithms
- search algorithms
- no subset selection of search algorithm portfolio
- no evaluation of different search problem encodings
- limited evaluation of feature selection
Results
Conclusions
- most kinds of Machine Learning perform poorly
- Support Vector Machines have best overall performance
- statistical relational learning promising
Questions?
Thank you!