It is shown in an abduction framework based on a general logic program [Eshghi 89, Kakas 92] that an abductive explanation for a given sentence might be provided by negation as failure when the negation of the sentence initiates an SLD resolution deduction. However, possibly non-ground explanations have never been expected. It is mainly because the SLDNF resolution (SLD resolution with negation as failure) is often restricted to keep a safe rule. In addition, only a case as a pragmatic example is to be extracted as an explanation for a query. In this paper, adopting non-safe SLDNF resolutions, we have an abduction framework in which a maximally general extension of abducibles is defined in terms of generality regarding the extensions. A general stable model semantics is firstly presented, to demonstrate the relation between the semantics of a program and the maximally general abductive extension, in the domain containing variables. A general stable model may denote a maximally general abductive extension which satisfies a constraint in abduction framework, while if a maximally general abductive extension satisfies the constraint then it may define a general stable model. We next prove that a non-safe SLDNF resolution is sound as an abductive proof procedure with respect to the maximally general extension of abducibles.

This paper proposes a view that AI can be a discipline to support any systems integration in any domain. The AI technique can be effective especially for requirement articulation for systems integration. However, so called knowledge-acquistion bottleneck has prevented AI from being widely used in systems integration. We propose to combine concept-formation aid systems with the knowledge-processing system. This combination can work as a fundamental system to support any systems integeration.

We propose a new framework for solving problems based on equivalent transformation of logic programs, where equivalent transformation is defined as changing programs preserving their declarative semantics. In the conventional logic programming, we (1) represent the knowledge in the given problem in terms of a logic program, (2) formalize the given problem as proving the given query from the knowledge, and (3) solve it in use of inference rules, such as resolution. In contrast, our new method does not use inference but equivalent transformation of logic programs. In the new method we (1) represent the knowledge together with the given query in the given problem as a logic program, (2) formalize the given problem as finding its equivalent logic program in a certain form, and (3) solve it by equivalent transformation of logic programs using equivalent transformation rules. Many problems, including the kind of problems which Prolog solves, are formalized and solved in the new method. The computational framework given here is called the Rule Based Equivalent Transformation (RBET). The validity of computation is strictly guaranteed by use of equivalent transformation rules, even if extralogical predicates are used in rules. Transformation rules are more expressive than Horn clauses in Prolog. RBET has the flexibility in the applications order of equivalent transformation rules, which makes efficient computation possible by the control of rule application based on rule preference.

In this paper, novel machine learning problem that handles clustering and a solution for this problem are described. The clustering is a method to divide a given set of individuals into clusters that are subsets having properties "internal cohesion" and "external isolation". The clustering is often used to get partition for a set of individuals desired to be fitted to some purpose. In such case, it is hard to define such desired partition by means of the properties internal cohesion and external isolation explicitly. But, it is usually easy to show desired partition itself. Therefore, by learning from examples that are pairs of a set of individuals and desired partition for the set, it is desirable to acquire a criterion that is used to get desired partition for any unknown set. In this paper, such a learning method is proposed. The method is different from ordinal "learning from examples". So we call it "Learning from Cluster Examples". In our prior work, the learning method is applied to a problem to divide a logic diagram image. The experimental results show that the proposed method learns more desired partitions than the ones in our prior work. This method is extended to be so general that is applied to logic diagram understanding but also dot pattern clustering or problems in other fields.

It is important for a student to understand that he/she made an error in problem solving in order to correct and prevent it from being repeated. We previously proposed a framework of a simulation which reflects an error a student made in solving a mechanics problem. Irregular and unnatural behavior of mechanics objects rcflecting an error helps a student understand that his/her solution is erroneous, because it visualizes what would happen based on his/her erroneous solution. We call the simulation "Error-Based Simulation (EBS)" and such a visualization "Error-Visualization". We have implemented a generator of EBS (EBS-generator) and evaluated the effectiveness of EBS on some examples of erroneous solution through an experiment. However, EBS isn't always effective for Error-Visualization. When EBS has only a quantitative difference from a normal simulation (NS), it isn't effective for a student to understand an error. EBS should have a qualitative difference from NS to be effective. Therefore, it is very important to diagnose the difference between EBS and NS in order to make use of EBS effectively for Error-Visualization. In this paper, we propose a framework for managing EBS and a method of its implementation based on qualitative reasoning techniques. The framework is based on an assumption that in order for EBS to be effective for Error Visualization, it should have a qualitative difference from NS in an object's velocity or in the ratio of an object's velocity's change to a parameter's change. The module which manages EBS is called EBS-manager. Its error management procedure consists of two phases. In Phase 1, by using qualitative simulation, behaviors of EBS and NS are predicted and then compared with each other. When a qualitative difference is found, EBS-manager judges that the EBS is effective. When a qualitative difference cannot be found, it proceeds to Phase 2. In Phase 2, by using comparative analysis, EBS-manager tries to find a parameter of which perturbation causes a qualitative difference between FBS and NS. When such a parameter is found, EBS-manager judges that the EBS with the perturbation is effective. When such a parameter cannot be found, EBS-manager judges that EBS isn't effective. We have implemented the EBS-manager and evaluated its effectiveness through an experiment. In this paper, we also discuss the result and outline our future work.

This paper presents an algorithm for unsupervised inductive learning from discrete probability distributions. Few algorithms have directly dealt with probability distributions. The procedures are as follows: 1. Find a probability distribution corresponding to a proposition of classical logic using maximum liklihood method. 2. Transform the probability distribution to a proposition based on the principle of indifference. 3. Reduce the proposition. The principle of indifference states that a probability distribution is uniform when we have no information. Using this principle, the propositions of classical logic can be corresponded to some probability distributions. The algorithm is applied to a real data. The result shows that the algorithm works well. The result is also compared with a method of multivariate analysis.

This paper presents an algorithm for discovering exceptional knowledge from databases. Exceptional knowledge, which is defined as an exception to a general fact, exhibits unexpectedness and is sometimes extremely useful in spite of its obscurity. Previous discovery approaches for this type of knowledge employ either background knowledge or domain-specific criteria for evaluating the possible usefulness, i.e. the interestingness of the knowledge extracted from a database. It has been pointed out, however, that the use of background knowledge can cause overlooking of useful knowledge. Furthermore, it is difficult to find such criteria in some domains. In order to circumvent these difficulties, we propose an information-theoretic approach in which we obtain exceptional knowledge associated with general knowledge in the form of a rule pair using a depth-first search method. The product of the ACEs (Average Compressed Entropies) of the rule pair is introduced as the criterion for evaluating the interestingness of exceptional knowledge. The inefficiency of depth-first search is alleviated by a branch-and-bound method, which exploits the upper-bound for the product of the ACEs. MEPRO (database Miner based on the average compressed Entropy PROduct criterion), which is a knowledge discovery system based on our approach, has been validated using the benchmark databases in the machine learning community.

As communication protocol consists of complex configurations of processes and has strict timing constraints, there have been many researches about formal verification including timing verification. We have reported the specification and verification method by timed automata. Using this method, we have verified the properties of fairness and regularity by language inclusion algorithm, but the verification costs are large. In this paper, we propose the effective timing verification method as follows. (1) We verify communication protocol both by the finite set of strongly connected components of specification and by checking timing constraints. (2) We verify communication protocol both by image computation based on BDD (Binary Decision Diagram) and by checking timing constraints. We have developed the verification system based on this method. This system can avoid state explosion problem.

Reinforcement learning is a framework in which an autonomous agent optimizes its bahavior by progressively improving its performance based on given rewards from the environment. Although several fruitful achievement has been made for the purpose of single-agent-adaptation by this framework, they are not applicable for multiple agents. To learn cooperatively, a new idea of reinforcement learning for multiple agents is needed. This paper describes a new method called Cooperative Reinforcement Learning with Spontaneous Mimetism where multiple agents in the environment learn cooperatively. First, we discuss two major problems of mimetism; when and whom to imitate. Next we compare Simple Mimetism where an agent always imitates on finding another agent in its neighborhood with simple reinforcement learning. To take advantages of both methods, we propose Adaptive Mimetism that adapts learning mode with balancing reinforcement learning and mimetism probabilistically by adjusting mimetism rate according to the situation. Finally, we show the merits of our method by the results of the simulation on the transportation problem in which several robots transport loads in the factory.