|
Prédictions contrôlées en apprentissage automatique. Alexander Gammerman, Vladimir Vovk. La revue MODULAD, numéro 42, 2010.
Mots clés
apprentissage automatique, prédicteurs conformes, complexité de Kolmogorov, approche bayésienne, étrangeté d’une prédiction
Abstract
Conformer predictors approach seems to be new and powerful. Its main advantage is that it is nonparametric and based only on the i.i.d. assumption. In comparison to the Bayesian approach, no prior distribution is used. The main theoretical result is the proof of validity of proposed conformal predictors. The second result is that asymptotically the relative number of cases when the real output value is within confidence interval converges to the average value of conformal predictors. The proposed technique is now applied to a large variety of practical problems. Two drawbacks of the approach are still mentioned in this discussion
Key words
Machine Learning, conformer predictor, Kolmogorov complexity, Bayesian approach, prediction strangeness
Article
|
