Topics that were discussed:

• What turnkey method provides the biggest gains, and under what circumstances?
• Are simple techniques (e.g. averaging) preferable to meta-learners (e.g., full learners that need to be trained and fine-tuned)?
• Is it even meaningful to ask these questions without domain knowledge?
• Can we explain why ensembles improve performance?
• Is it better to use repeated runs of one algorithm or of many different algorithms with different levels of performance?
• Should one use all available data in training the individual generalizers (each starting with different initial algorithm parameters)?
• Should one use spatial partitions of the data (e.g., mixtures of experts)?
• Should one use statistical partitions of the data (e.g., bagging)?
• Should one use partitions of the data based on previous performance (e.g., boosting)?
• Should one distort the spaces (e.g. ECOCs, adding noise to inputs of individual learners)?
• Should learners be trained/constructed differently if they'll be used in a turnkey technique than if they'd be stand alone generalizers?
• Are there advantages in actively trying to reduce the correlation among algorithms in a pool?
• How can the experimental gains due to turnkey algorithms be reconciled with the "No Free Lunch" (NFL) results? (e.g., what are the assumptions that lead to improvements?)

If you have any questions or comments, please email Kagan Tumer.

Return to the turnkey algorihtms main page.