By Koichi Furukawa, Keio University, Japan
The aim of this report is to introduce research activities related to Inductive logic programming (ILP) in Japan. ILP research in Japan has been conducted actively since it appeared in Europe in 1991. ICOT, the Institute for New Generation Computer Systems, hosted the second International Workshop on Inductive Logic Programming in 1992, co-chaired by Stephen Muggleton and myself. Since then, several research groups in Japan studied on ILP in various aspects including basic theory, system development and applications. In 1998, Setsuo
Arikawa from Kyushu University launched the Discovery Science Project sponsored by Grant-in-Aid for Scientific Research on Priority Area from the Ministry of Education, Science, Sports and Culture (MESSC) of Japan. This is a three-year project started from 1998 that aims to (1) develop new methods for knowledge discovery, (2) install network environments for knowledge discovery, and (3) establish Discovery Science as a new area of computer science. One of the research groups, headed by Taisuke Sato from Tokyo Institute of Technology, is devoted to promote research on discovery science by reasoning and by ILP.
2. ILP research in Discovery Science Project
There are six groups, which conduct ILP researches in the project. They are my group at Keio University, Taisuke Sato's group at Tokyo Institute of Technology, Hiroki Arimura's group at Kyushu University, Makoto Haraguchi's group and Akihiro Yamamoto's group at Hokkaido University, and Fumio Mizoguchi's group at Science University of Tokyo. My group is conducting ILP researches from three aspects: theory, system and application. Our theoretical work focuses on the completion of inverse entailment. Furukawa found a sufficient condition that inverse entailment computes the most specific hypothesis (MSH) which is subsumed by any sound hypotheses. The work is deeply related Yamamoto's work. Furukawa further investigated an extension of the MSH, which is sound and complete. System research in our group is to build a system to bridge ILP to RDB by appropriate SQL interface. As an application research of ILP, we are building a simulator of children's vocabulary acquisition process in terms of ILP. We adopted the constraint theory proposed by cognitive science researchers. We succeeded in demonstrating a simple vocabulary acquisition session in the domain of animal classification.
Taisuke Sato's group is devoted to integrate probabilistic behaviors to logic programming and learn best distribution based on EM algorithm.
Hiroki Arimura's group is working on giving PAC learnability condition for ILP. There have been no researches to formulate PAC learnability in ILP and therefore their work really has brought a break through in this domain.
Makoto Haraguchi's group tries to apply ILP to program synthesis in the form of a two-dimensional diagram called Intelligent PAD. They are also working on background knowledge with hierarchical structure.
Finally Yamamoto's group is working, as mentioned earlier, on a theoretical issue to investigate the completeness of inverse entailment. They are developing their own ILP system based on sound operations called Bottom Generalization Method, which is similar to Golem, by Muggleton.
Fumio Mizoguchi's group is developing an ILP system called GKS, which incorporates constraint logic programming in ILP, and induces rules with constraints. They succeeded in applying their system to house designing domain and extract some general design rules from give examples.
Research activities on ILP are becoming very active in Japan. The research spectrum is very broad and we expect plenty of fruitful results will be produced from these researches. Discovery Science project is organizing an International Conference on Discovery Science 1999 from December 6 to 8, 1999, at Waseda University, Tokyo Japan. Those who are interested in participating the conference should refer the URL http://www.i.kyushu-u.ac.jp/ds99.