Núria Macià, Jaume Bacardit, and Ester Bernadó-Mansilla

                                 nmacia@salle.url.edu
                     jaume.bacardit@nottingham.ac.uk
                                  esterb@salle.url.edu
Methodology for learners’ assessment
The UCI repository. 134 classification problems



Source: GECCO’11 Proceedings
Algorithm refinement        Knowledge extraction




              Behind the experiments
Standard comparison



Source: Jaume Bacardit, Edmund K. Burke, and Natalio Krasnogor. Improving the scalability of rule-based evolutionary learning. (2009)
Taxonomy of problems
Complexity coverage
• How many data sets should we use in the experiments?
• Which ones? Synthetic data sets? Real-world problems?
  Both?
• Is the UCI our best sample?
• How should we select our referenced learners?
• Should we keep performing comparisons over an arbitrary
  set of problems?
• How can we characterise problems?
• More coverage? Benchmarks?
• Should we tackle one problem at a time?
• …
                      Questions. Answers?
Núria Macià, Jaume Bacardit, and Ester Bernadó-Mansilla

                                 nmacia@salle.url.edu
                     jaume.bacardit@nottingham.ac.uk
                                  esterb@salle.url.edu

Testing learning classifier systems

  • 1.
    Núria Macià, JaumeBacardit, and Ester Bernadó-Mansilla nmacia@salle.url.edu jaume.bacardit@nottingham.ac.uk esterb@salle.url.edu
  • 2.
  • 4.
    The UCI repository.134 classification problems Source: GECCO’11 Proceedings
  • 5.
    Algorithm refinement Knowledge extraction Behind the experiments
  • 6.
    Standard comparison Source: JaumeBacardit, Edmund K. Burke, and Natalio Krasnogor. Improving the scalability of rule-based evolutionary learning. (2009)
  • 7.
  • 8.
  • 9.
    • How manydata sets should we use in the experiments? • Which ones? Synthetic data sets? Real-world problems? Both? • Is the UCI our best sample? • How should we select our referenced learners? • Should we keep performing comparisons over an arbitrary set of problems? • How can we characterise problems? • More coverage? Benchmarks? • Should we tackle one problem at a time? • … Questions. Answers?
  • 10.
    Núria Macià, JaumeBacardit, and Ester Bernadó-Mansilla nmacia@salle.url.edu jaume.bacardit@nottingham.ac.uk esterb@salle.url.edu