Reinforcement Learning for Rule Generation
The algorithm extracts propositional rules from a labeled data set. The constituent parts of a rule are the features of the labeled data-set, each accompanied by an appropriate interval of activation and a label denoting the class. Initially, the input space is partitioned using tiles. The algorithm tries to compose the largest possible orthogonal intervals out of tiles. After the creation of intervals for each feature the rule receives credit for its classification ability. This credit will be used to improve the rule. We have obtained encouraging results on 5 different classification problems: the iris data set, the concentric data, the four gaussians, the pima-indians set and the image segmentation data set.
KeywordsHide Layer Input Space Rule Extraction Classification Ability Eligibility Trace
Unable to display preview. Download preview PDF.
- R. S. Sutton A. G. Barto, Reinforcement Learning. The MIT Press, 1998.Google Scholar
- V. Cherkassky, Learning from Data, ch. 6. John Wiley & Sons, INC., 1998.Google Scholar
- R. S. Sutton, “Implementation details of the TD(⋋) procedure for the case of vector predictors and backpropagation,” Tech. Rep. 87-509.1, GTE Laboratories Incorporated, Aug 1989.Google Scholar
- “ftp.ics.uci.edu/pub/machine-learning-data bases, ftp.dice.ucl.ac.be/pub/neural-nets/elena/databases/artificial/concentric/.”Google Scholar
- R. Munos A. Moore, “Barycentric interpolators for continous space & time reinforcement learning” Neural Information Processing Systems, 1998.Google Scholar
- M. Sato S. Ishii, “Reinforcement learning based on on-line em algorithm” Advances in Neural Information Processing Systems, vol. 11, 1999.Google Scholar