TY - GEN

T1 - Mutual information gaining algorithm and its relation to PAC-learning algorithm

AU - Takimoto, Eiji

AU - Tajika, Ichiro

AU - Maruoka, Akira

PY - 1994/1/1

Y1 - 1994/1/1

N2 - In this paper, the mutual information between a target concept and a hypothesis is used to measure the goodness of the hypothesis rather than the accuracy, and a notion of mutual information gaining (Mi-gaining) algorithms is introduced. In particular, strong and weak Mi-gaining algorithms are defined depending on the amount of information acquired, and their relation to strong and weak PAC-learning algorithms are investigated. It is shown that although a strong Mi-gaining algorithm is equivalent to a strong PAC-learning algorithm, a weak MI- gaining algorithm does not necessarily imply a weak PAC-learning algorithm, and vice versa. Moreover, a general boosting scheme for weak Mi-gaining algorithms is given. That is, any weak Mi-gaining algorithm can be used to build a strong one. Since a strong Mi-gaining algorithm is also a strong PAC-learning algorithm, the result can be viewed to give a sufficient condition for a class of algorithms to be boosted into strong learning algorithms.

AB - In this paper, the mutual information between a target concept and a hypothesis is used to measure the goodness of the hypothesis rather than the accuracy, and a notion of mutual information gaining (Mi-gaining) algorithms is introduced. In particular, strong and weak Mi-gaining algorithms are defined depending on the amount of information acquired, and their relation to strong and weak PAC-learning algorithms are investigated. It is shown that although a strong Mi-gaining algorithm is equivalent to a strong PAC-learning algorithm, a weak MI- gaining algorithm does not necessarily imply a weak PAC-learning algorithm, and vice versa. Moreover, a general boosting scheme for weak Mi-gaining algorithms is given. That is, any weak Mi-gaining algorithm can be used to build a strong one. Since a strong Mi-gaining algorithm is also a strong PAC-learning algorithm, the result can be viewed to give a sufficient condition for a class of algorithms to be boosted into strong learning algorithms.

UR - http://www.scopus.com/inward/record.url?scp=0012417189&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=0012417189&partnerID=8YFLogxK

M3 - Conference contribution

AN - SCOPUS:0012417189

SN - 9783540585206

T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)

SP - 547

EP - 559

BT - Algorithmic Learning Theory - 4th International Workshop on Analogical and Inductive Inference, AII 1994 and 5th International Workshop on Algorithmic Learning Theory, ALT 1994, Proceedings

A2 - Arikawa, Setsuo

A2 - Jantke, Klaus P.

PB - Springer Verlag

T2 - 4th International Workshop on Analogical and Inductive Inference, AII 1994 and 5th International Workshop on Algorithmic Learning Theory, ALT 1994

Y2 - 9 October 1994 through 14 October 1994

ER -