This paper presents a method to learn nonmonotonic rules with exceptions from positive/negative examples and background knowledge in Inductive Logic Programming. We adopt extended logic programs as the form of programs to be learned, in which two kinds of negation, negation as failure and classical negation, are effectively used in the presence of incomplete information. We implemented the learning system LELP. LELP first generates candidate rules from positive examples and background knowledge using ordinary ILP techniques. Default rules with exceptions are then generated as rules with negation as failure using the OWS algorithm. Exceptions are identified from negative examples and are then generalized to default cancellation rules. In LELP, default rules can be produced for both positive and negative consequents. Then, if some instances are possibly classified as both positive and negative, nondeterministic rules are constructed to resolve the inconsistency by getting multiple answer sets. Moreover, hierarchical defaults can also be learned by recursively calling the exception identification algorithm.

In this paper, we propose the dynamic position arrangements of the soccer agents for realizing the autonomous team play. In addition, we adopt the genetic algorithms for obtaining the adaptive position arrangements to defeat an opponent team. In real soccer game, the formations of players are important factors to defeat opponent teams. The formations are the arrangements of the player's positions. The formations will be changed according to the ball position. Namely, each player moves around the given positions to realize the effective team play. Therefore, we use the analogy of the adaptive formation for constructing the autonomous team play with the soccer agents. That is, the dynamic position arrangement method is proposed. The agents can perform several behaviors (pass, shoot, search the ball, approach the ball, go back to own home position) around one fixed home position. The agents can take several home positions in each game, and can dynamically change current home position in accordance with the situations in the game. The agent can move around the home positions and can perform the suitable behavior for current situation if the agents are given the appropriate home positions. To search the appropriate home positions, we adopt the genetic algorithms as robust optimization methods. To apply it to the genetic algorithms, we must decide what function we should use in order to evaluate the arrangement of the positions. It is very difficult problem to judge the effectiveness of the position arrangements. Because even if the team having the effective position arrangement, the team can't always win the games. Namely, the result of the games includes many noises. So, we assume the evaluation function based the behavior as follows; In the team level, if the total occurrence of the pass behavior and the shoot behavior in the game are increased than others, the position arrangements may be effective than others even if these behaviors didn't succeed. In the individual level, each agent should dash to catch the ball immediately when the ball is closed to the agent. While, the agent should go back own home position when the ball is far from the agent. Of course, the number of the goals which the team gets is added to evaluate the arrangement of positions. Thus, we suggest the behavior based evaluation function which consists of the frequency of each agent's behavior and the score. We will show the experimental results applied the proposed methods on the soccer server.

A system that interactively explains a machine and its installation by utilizing linguistic information and visual information was built. This system has two frameworks in which interactive multimodal explanation can be generated. The first framework uses formatted text and pictures. It is interactive as it allows the users to make follow-up questions and replies using spoken language with pointing gestures. The second framework uses spoken explanation coordinated with pointing gestures and animation. That explanation is a reproduction of instruction dialogues made by experts in a face-to-face situation. This paper describes the mechanisms for explanation generation in this interactive multimodal explanation system. The mechanisms needed for achieving the interactive features of explanation such as accepting follow-up questions, and the way of handling the temporality of explanation caused by the use of temporal media such as spoken language are explained in detail. In addition, through the comparison of the two frameworks of multimodal explanation and their generation mechanisms, an appropriate organization of processing modules and knowledge sources for multimodal explanation generation is proposed. The description of the object explained and means available for explanation such as an utterance realizer can and should be shared between frameworks under the consideration that they are used in different ways. Explanation generation mechanisms themselves, however, should be designed specifically for each framework. The difficulties of sharing those components come from two facts. First, there is divergence in the rhetorical devices that can be utilized and the factors that need to be considered in making explanations easy to understand. Second, the way of making explanation interactive differs in principle in each explanation framework.

We propose a method of compiling circumscription into extended logic programs which is widely applicable to a class of parallel circumscription as well as a class of prioritized circumscription. In this paper, we show theoretically that circumscription whose theory contains both the domain closure axiom and the uniqueness of names axioms and does not contain function symbols can be compiled into an extended logic program II, so that, whether a ground literal is provable from circumscription or not, can be evaluated by deciding whether the literal is true in all answer sets of II, which can be computed by running II under the existing logic programming interpreter. The proofs of all theorems w.r.t. the method are also presented.

Genetic algorithm (GA) is an effective method to solve combinatorial optimization problems. The efficiency depends on how genes expressed. Although gene expression using finite state automata has already been proposed, it is not always suitable for genetic operation. We propose an gene expression using n-BDD, which lets genetic operations work efficiently. We show a simulation of Quasi-Ecosystem including herbivores, carnivores and plants as an experimental model to confirm the efficiency of our method compared with systems using finite state automata and classifier systems. Our experiments tell superiority of the n-BDD expression at the adjustment speed of animals' action strategy. In the Quasi-Ecosystem, strategies of animals' behavior are evolved in the genetic expressions. We also observed food chain in the environment, in which quick evolution of animals' behavior is necessary, using n-BDD expression.

This paper presents a method of knowledge-base reconstruction (KBR) by means of changing the viewpoint of exceptions within the knowledge. The viewpoint means how the exceptions of a class should be decided when we represent the general property of the class by incomplete knowledge which allows exceptions. If there are multiple viewpoints, the different incomplete knowledge is considered according to the viewpoint, and each of them has the different exceptions. Therefore the change of the viewpoint enables us to change both the quantity of the exceptions and the quantity of the rules which are related with these exceptions. By focusing on the viewpoint to reduce these rules, we can make a compact knowledge-base (KB). We realize this KBR by six primary operations classified into the following three types: 1) specialization and generalization of an incomplete rule which represents the incomplete knowledge, 2) addition and deletion of an incomplete rule, and 3) changing the property of a rule. Since the KB must describe the same world before and after this KBR, it should be guaranteed that the application of these operations does not change any conclusions derived from the KB. After proving the intangibility of the conclusions, we propose an algorithm for changing the viewpoint of exceptions.

This paper proposes a thinking support system for idea inspiration. The system visualizes data extracted from texts on a network and represents them in two dimensional metric space on a spring model. Up to now many tools which give a person a trigger which extend idea have been proposed. The problem of these system is that these tools are not equipped with powerful interactive interface in order that user's point of view can reflect to the system. This interactive interface is important, because it makes user enable to get more various points of view to reach new idea via this interface. In this system, through extending spring model method, an interactive interface is provided. The system visualizes similarity relations between texts and keywords in two dimensional metric space. Furthermore a user can request this system to add, move and delete each object. At this time it can gradually be arranged after the request. The system can show a process when it visualizes a new arrangement. Finally through experiments, the proposed system is shown to be effective for giving a person a trigger to extend idea. In future we try to merge other thinking support systems which support brain storming and kJ method.

For a computer system to generate expressive music rendition, especially by applying performance rules obtained empirically, it has to find out portions of a musical score where qualitatively expressed performance rules are applied. It is often the case that those portions are described in terms of music analysis. Daphne is a computer-assisted music analysis system on which users can specify or inquire the more abstract information including structure and metre which are implied in a musical score. Information obtained through Daphne is shared not only among computer music systems, but also between human and systems and by several people in order to overcome the ambiguity of music knowledge for improving computer music systems. For the purpose of information sharing, the contents-oriented ontology can be a powerful framework where concepts are clarified into language and the implied primary knowledge is brought to implementation. In this paper, the expectation to music ontology for building computer music systems as well as the design and implementation of Daphne is described.

In this paper we discuss an approach to the inference of recursive equational programs from non-recursive equations for the target functions. We call the research area Inductive Equational Programming (IEP). We firstly give a definition of IEP problems. Then we give three inference methods for inferring target programs from given concrete examples under some background knowledge. The inference methods are inverse reduction, generalization, and finding subsidiary functions which are not given in examples. We show the soundness of these inference methods. We also compare them with other methods which have already been developed for inference of recursive functions.

Abduction is a form of nonmonotonic reasoning. It is constituted by an abduction framework (P^*, Ab^*, 1^*), where P^* is an abductive program, Ab^* is a set of abducible atoms and I^* is an integrity constraint. Its declarative semantics is defined by abductive extensions P^* ∪Δ such that Δ ⊆ Ab^* and P^* ∪ Δ satisfies I^*. Eshghi and Kowalski (1989) have given the abduction framework whose declarative semdntics is equal to (2-valued) stable models and the abductive proof procedure to obtain an abductive explanation for a query. But this procedure is not sound with respect to this semantics. Dung (1991) has shown proper semantics, e.g. preferred extension, which generalizes other declarative semantics for normal logic programs. The abductive proof procedure is sound with respect to Dung's semantics. In this paper, we establish an abduction framework which is based on 3-valued logic. It is realized by dealing with the abductive adjustment which is a set of abducible atoms interpreted undefined. In 3-valued abduction framework, its declarative semantics is defined by 3-valued abductive extensions P^* ∪ Γ^u ∪ Δ. While Dung's semantics is in terms of the notion of acceptability in addition to the integrity constraint, our semantics only relies on the integrity constraint. This paper shows that their semantics are equivalent. We also present the relations between our framework and the alternating fixpoint by van Gelder (1993). Finally, we have the relations amont Dung's abduction framework, our 3-valued abduction framework and alternating fixpoint.

Classifier systems are now viewed disappointing because of their problems such as the rule strength vs rule set performance problem and the credit assignment problem. In order to solve the problems, we have developed a hybrid classifier system: GLS (Generalization Learning System). In designing GLS, we view CSs as model free learning in POMDPs and take a hybrid approach to finding the best generalization, given the total number of rules. GLS uses the policy improvement procedure by Jackal et al. for the optimal stochastic policy when a set of rule conditions is given. GLS uses GA to search for the best set of rule conditions.

In this paper, we study automated theorem proving based on Model Elimination (ME) tableau and genetic programming. We apply several GP-based search methods to propositional ME tableau calculus, and evaluate the performance of them experimentally.