Paper 3 AlgoSelect

Question 1

What are the prerequisites for algorithm selection?

Answer 1

Availability of large problem collections
Existance of large number of algorithms
Performance metrics for algo on problem
Existance of suitable features to characterize properties of instances

Question 2

What does combining the prerequisits allow?

Answer 2

Creation of comprehensize set of meta-data for learning algo performance

Question 3

4 essential components of the model/framework

Answer 3

Problem space P: set of instances of problem class
Feature space F: measurable characteristics of feautres extraced from P
Algorith space A: set of all considered algorithms
Performance space Y: mapping of algorithm to performance metrics

Question 4

Rice algo select

Answer 4

For a problem x, extract features f(x), select a=S(f(x)), check performance of a on x: y(a(x)), maximise performance.

Question 5

Aha [1992]

Answer 5

Introduce rule based learning of algo select

Question 6

No free lunch

Answer 6

Across all problems, the performance of all algorithms is equally bad

Question 7

What did StatLog add?

Answer 7

Used a lot of features to create rules, found that sometimes calculating features is more work than running simple algorithms. (Could performance be predicted?) (Landmarking instead of features)

Question 8

METAL

Answer 8

Added model selection and method combination approaches