Inductive Logic ProgrammingやAbductive Logic Programmingに関する最新の研究 成果の報告が相次いだ。海外招 待講演者も含み、この分野の専門家が多数同席したの で、質疑も活発で、大いに議論が盛り上がった。研究発表２１件 のうち英語での発表 １４件（うち２名は海外からの参加）、日本語の発表は７件であった。従って、２５ 件１８件が英 語の発表で、巧みなジョークの飛び交うバンケットもあり実質ミニ国際 会議であった。このような企画を提案され、海 外招待講演者４名のスポンサーでも あった慶応大学古川教授に感謝する。
・テーマ ： 一般演題及び小特集「Inductive Logic Programming」
・日時 ： 1996年7月31日(水)14:00 〜 8月2日（金）12:30
・場所 ： 北海道大学百年記念館
・担当幹事： 國藤 進 E-mail: email@example.com
July 31 (Wed)
14:00-14:10 Opening Session Chair S. Kunifuji (JAIST) Opening Address K. Furukawa (Keio University)
14:10-15:30 Session 1 Chair K. Furukawa (Keio University)
1) Invited talk from abroad (1) J. A. Robinson(University of Syracuse):
Acquisition and Deployment of Motor Skills in Musical Performance
Abstract: Musical performance depends on motor skill modules acquired by repetitive training ("practice"). These are invoked by higher level knowledge dealing with the musical architecture. Such higher level knowledge can be transferred by teaching and reading. The lower level motor skills cannot. The best performers often do not know how they play certain passages. Machine learning could help to improve human ability to understand and master performance difficulties.
2) Yuji Ishikawa (NTT Data Communications Systems Corp.) and
Makoto Haraguchi (Hokkaido University):
A Theory of Knowledge Revision using Multiple Generalizations
Abstract: This paper presents a new framework for revising knowledge in terms of a subset of F-logic. Since every rule in F-logic denotes a complex body of knowledge about objects, our revision method is designed so that it operates on object molecules as possible as it can. It never depends on atomic formulas obtained by translating F-logic into the standard first-order formalism. Our revision task is supposed to be invoked by both positive and negative observations about objects. So we consider a new technique called a multiple generalization of rules. By the multiplicity, we can compute a set of generalized rules not covering negative observations about objects.
15:50-17:50 Session 2 Chair S. Kunifuji (JAIST)
3) 有川 節夫、橋本 信幸（九州大学）: 数値データからの微分方程式の学習
概要： 誤差を含んだ数値データから、微分値を閉区間の形で推定することにより、 微分方程式を学習する手法を提案する。
4) 有馬 淳（富士通情報研）： 有機的相同体計算モデル -- 文脈に厚いシステムにむけて
概要： 多様な環境に相同体が自己組織化し相互作用することで適合する 頑健なシステムのモデルを提案する。
5) 三浦 純,白井 良明(大阪大学):
概要: プラニング時間とプランの質のトレードオフを考慮して移動ロボットの視覚 と行動のプラニングを行う．
6) 岡 雄三(工学院大学), 兒玉光弘(桐蔭学園横浜大学):
August 1 (Thu)
9:00-10:10 Session 3 Chair S. Kunifuji (Hokkaido University)
7) Invited talk from abroad (2) Randy Goebel(University of Alberta):
If ILP leads, will database mining follow?
Abstract: The increasing popularity of inductive logic programming (ILP) has provided one clear demonstration that machine learning has become practical. A more general area in which induction has a role is knowledge discovery in databases (KDD). There too induction has a role, but more of the current developments are based on the creation of abstraction rules, guided by the use of explicit concept hierarchies and ``hypothesis'' rankings. We examine some of the directions in KDD, with the goal of identifying ILP research that can gracefully lead KDD to improved methods.
8) T. Kanai and S. Kunifuji(JAIST): Abductive Inductive Logic Programming
Abstrct: We propose an integrated framework of Abductive Logic Programming and Inductive Logic Programming. We show that the integrated framework can deal with the problems of induction using incomplete background knowledge.
10:30-12:00 Session 4 Chair: M. Haraguchi (Hokkaido University)
9) Fumio Mizoguchi, Hayato Ohwada and Makiko Daidoji(Science University of Tokyo):
Learning Design Rules for Spatial Layout based on Inductive Logic Programming
Abstract: One of the promising tasks in knowledge-based design synthesis is automatically extracting useful rules from past design experience. This paper demonstrates how inductive logic programming (ILP) is applied to automated acquisition of rules and/or constraints for spatial layout such as Floor Planning. Such new ILP applications need extensions of current ILP framework in two directions. First, rules with spatial constraints are learned from numerical data, and are executed as a constraint logic program. Second, multiple targets are handled to identify the ordering among a number of spatial relationships, and to construct a hierarchy of learned hypotheses. These facilities are incorporated into our ILP system GKS, which progressively produces constrained rules with maximum predictive values. It is also demonstrated that a set of learned rules dramatically reduce search space of combinatorial problems and enhance the problem-solving performance of the original Floor Planning System.
10) Fumio Mizoguchi and Hayato Ohwada(Science University of Tokyo):
Personalized E-mail Agent using Inductive Logic Programming
Abstract: This paper outlines an Inductive Logic Programming approach to designing personalized E-mail agent (PMAIL). In this approach, personality is acquired from user's mail characterizations such as category, priority and preference. After saving user's behaviors on mail handling, PMAIL progressively produces logical rules that characterize new mails and are used to provide users with ``prior knowledge'' for looking over the mails. Since such rules are comprehensive and readable, PMAIL has the advantage of explicit interaction for personal assistance. The competence of PMAIL is due to the performance of induced rules, and therefore we undertake an empirical study by introducing a variety of performance measures such as predictive accuracy. In this study, we show how PMAIL's learning parameters affect the performance and how induced rules are applied to enhance the performance task. The present study indicates the expressive and potential power of Inductive Logic Programming toward realizing personalized systems.
11) Y. Murakawa(JAIST), T. Washio(Osaka University) and S. Kunifuji(JAIST):
A learning mechanism for the selection of hypotheses on
Abductive Reasoning and it's application.
Abstract: We propose a learning mechanism that is to learn how to select hypotheses from a set of abducibles (possible hypotheses) on abductive reasoning. Abductive reasoning is to infer an explanation of why observations could have occurred. In abuduction this explanation is called a hypothesis and it is selected from a set of the given possible hypotheses. This selection follows the probabilistic heuristics (ME (Minimal Explanation) standard and LPE (Least Presumptive Explanation) standard). This learning mechanism is to learn preferentially propositions or rules that are selected by the probabilistic heuristics. So the explosion and the ambiguity of learning knowledge are reduced. This is an advantage of our proposing method comparing with inductive learning. And we show an application of our learning mechanism for WWW.
12:00- 13:20 Lunch
13:20-14:50 Session 5 Chair K. Inoue (Toyohashi Univ. of Technplogy)
12) Takashi Ishikawa (Kisarazu National College of Technology) and
Takao Terano (The University of Tsukuba)：
Concept-Description Analogy based on Concept Similarities
Abstract: This paper proposes an algorithm for analogical reasoning of concept-descriptions in a framework of logic programming. The algorithm learns the description of a predicate similar to a given predicate from a few examples of the target predicate using similarities defined with taxonomic information represented by first-order predicate logic.
13) Hiroshi Fujita (Kyushu University), Naoki Yagi (Fujitsu),
Tomonobu Ozaki, Koichi Furukawa ( Keio University):
A new design and implementation of PROGOL by bottom-up computation
Abstract: We present a new design and implementation of PROGOL by using MGTP (Model Generation Theorem Prover) technology. Specifically, we apply the technology to the computation of the most specific hypothesis given a positive example and background knowledge. We also use the technology to calculate an evaluation function for obtaining the best hypothesis.
14) Keiko Shimazu(Fuji Xerox Corp.) and Koichi Furukawa (Keio University) :
Knowledge Discovery in Database by PROGOL
-- Design, Implementation and its Application to Expert System Building
Abstract: We introduce a framework for realizing Knowledge Discovery in Database by utilizing ILP technology. We study (1) how to design a target concept representation, (2) how to identify and define necessary background knowledge, (3) how to generate negative examples automatically, and (4) how to restrict the PROGOL search space to a finite domain. Then, we apply the framework to automatic knowledge acquisition of expertise in an e-mail classification expert system.
15:10-17:50 Session 6 Chair S. Tojo (JAIST)
15) Invited talk from abroad (3) Claude Sammut(University of South Wales):
"Inductive logic programming and multistrategy learning"
Abstrct: One of the most important contributions of ILP to machine learning is its ability to incorporate background knowledge into a learning task in a clean and easy manner. A special advantage of learning in the context of a Horn clause representation is that both background knowledge and learned concepts are executable as programs. Thus, background knowledge can be made quite complex. In this paper we discuss the implications of incorporating other learning algorithms as background knowledge into an ILP system. We claim that ILP provides an excellent framework for combining the benefits of different learning strategies.
16) Alipio Jorge and Pavel Brazdil(University of Porto):
Integrity Constraints in ILP using a Monte Carlo Approach
Abstract: This paper describes a Monte Carlo strategy to check the consistency of a logic program and a set of integrity contraints. The integration of this very efficient technique with ILP learning system SKILit is described.
17) Xiaolong Zhang and Masayuki Numao(Tokyo Institute of Technology):
On the utility of clause reuse in inductive logic programming
概要： 背景知識がインクリメンタルに追加された場合に,節を再利用しながら, 効率的に帰納学習を行なう手法について述べる。
18) Masayuki Numao and Yuko Wakatsuki(Tokyo Institute of Technology):
ILP on Decision Trees
概要： 命題記述の学習方式が有利だが，ごく一部で述語記述を要する応用を想定し， 決定木を拡張したILPを提案する．
19) Edison S. Gomi and M. Ishizuka (Univ. of Tokyo)：
A Method for Induction of Recursive Logic Programs
Abstract: Inversion of logical implication is an important problem in Inductive Logic Programming (ILP). For clauses, logical implication can be verified using the θ-subsumption relationship. However, θ-subsumption is incomplete with respect to implication. This problem occures for self-recursive clauses, and consequently, for recursive logic programs. This paper presents MiMFoS, an ILP system designed to induce recursive logic programs belonging to a restricted class named two-clause linear recursive closed ij-determinate logic programs. MiMFoS employs an extension of a method named forced simulation. In order to generate the initial hypothesis, it had been necessary to consult an oracle that classifies the positive examples as base or recursive cases. MiMFoS implements this oracle using an algorithm called 2-minimal multiple generalization (2-mmg) algorithm.
18:00 - 20:00 Banquet
August 2 (Fri)
9:00-10:40 Session 7 Chair M. Numao (Tokyo Institute of Technology)
20) Invited talk from abroad (4) Ashwin Srinivasan(Oxford University):
Feature construction with Inductive Logic Programming:
a study of quantitative predictions of chemical activity
aided by structural attributes
Abstract: This talk examines the use of ILP programs, not for obtaining theoriescomplete for the sample, but as a method of ``discovering'' new attributes. These could then be used by methods like linear regression, thus allowing for quantitative predictions and the ability to use structural information as background knowledge.
21) M. Oomori and S. Tojo(JAIST)： Reasoning on situated information
Abstrct: We propose a situation-based model on reasoning system with a sort hierarchy that provides natural inference steps for human beings. We show that sort hierarchy can be partially changed according to the progress of event sequence and that the reasoning using situated information works well for legal domain where temporal relations between affairs are important.
22) Paulo Azevedo（Universidade do Minho）： Semantic Subsumption for Magic Sets
Abstrct: In this paper we study the relationship between tabulation and goal-oriented bottom-up evaluation of logic programs. Differences emerge when one tries to identify features of one evaluation method in the other. We show that to obtain the same effect as tabulation in top-down evaluation, one has to perform a careful adornment in programs to be evaluated bottom-up. Furthermore we propose an efficient algorithm to perform subsumption checking over adorned magic facts.
10:40 - 11:00 Break
11:00 - 12:30 Session 8 Chair J. Arima (Fujitsu Lab.)
23) 秋葉 澄孝（電子技術総合研究所）: 節の生成方法について
24) 山本 英子,井上 克已 (豊橋技術科学大学): 線形導出に基づく結論発見プログラムの効率化
概要： 線形導出法の拡張であるSOL導出に基づき一階述語論理における 結論発見プログラムを実現し効率化を図った。
25) 井上 克已,中西 博一 (豊橋技術科学大学）： 失敗による否定を含む規則の自動生成
概要： 負例を例外とみなし、安定モデル意味論に従うような、否定を含む 常識規則を生成する手法について述べる。