There are two types of research styles in AI such as "Form-oriented" and "Content-oriented" and the former has dominated the AI research to date. Recently, however, the importance of "Content-oriented" research has been recognized. This is partly because several problems to solve require not only formal theories but also technologies dealing with content of knowledge. Although content-oriented research seems promising, which is partly justified the success of expert systems research, we have to admit that content-oriented research is suffering the following three difficulties: 1) The research tends to become ad-hoc, 2) we do not have any sophisticated technologies for enabling the research results to accumulate, and 3) we have no basic research. Without overcoming these shortcomings, we could not expect the growth of the research. The main objectives of this paper is to propose a new research field called "Ontology Engineer-ing" and to show it can be a basic research of content-oriented research and provide such technologies badly needed by it. We begin the paper by discussing what an ontology is. Although ontology is becoming popular within a community, it is not well understood in AI community in general. We carefully explain what an ontology is. To our knowledge, how to use an ontology is one of the crucial issues in ontology research. Therefore, we analyze the depth of the ontology use in seven levels as well as discussing what concrete advantages ontology can give in the real-world problem solving. The next topic is the classification of ontologies in which we introduce an idea of specification of problem solving context by identifying task ontology. We exemplify some content-oriented research including our recent work. Finally, on the basis of the discussion made thus far, we present some research agenda of ontology engineering which covers philosophy, knowledge representation, general and domain ontologies, standardization, electronic data/knowledge interchange, knowledge reuse/sharing, media integration, and so forth.

For a physical robot to acquire behaviors, it is important for it to learn in the physical environment. Since reinforcement learning requires large computation costs as well as a lot of time in the physical environment, most research has performed learning by simulation. However, this does not work well in the real world. Realizing reinforcement learning of a physical robot in a physical environment requires both an adaptation for the diversity of possible situations and a high-speed learning method that can learn from fewer trials. This paper describes cooperative reinforcement learning based on propagating the learned behaviors of a virtual agent to a physical robot in order to accelerate learning in a physical environment. The method consists of two parts: (1) preparation learning in a virtual environment to accelerate initial learning, which accounts for most of the learning cost ; and, (2) refinement learning in a physical environment by using the virtual learning results as an initial behavior set of a physical robot. Experimental results are given for a ball-pushing task with the physical robot and a virtual agent.

Existing inductive logic programming (ILP) systems learn only one target predicate from a set of background predicates and a set of positive and negative examples. Therefore, such ILP systems cannot deal with problems which involve several target predicates. This paper presents a multiple predicate learning (MPL) system with a new learning approach. MPL systems are usually prevented from learning target predicates when the constants appearing in the examples of one target predicate differ from those of the others, e.g., the Dutch flag problem is problematic for existing MPL systems. Our system, MPL-Core, has ability to tackle the case and efficiently learns from multiple predicate tasks. Core, a single predicate learning module, has a fast failure mechanism and can select refinement operators based on the learning task. By means of GPC, an efficient pruning method, Core effectively prunes unpromising branches in a search tree, making the search space more tractable. Furthermore, our algorithm uses the fast failure mechanism which gives it a distinct advantage over existing multiple predicate learning algorithms in terms of computational complexity. Experimental results show the learning efficiency and potential learning ability of MPL-Core. The effect of the fast failure mechanism is also shown by experimental comparisons on learning from both artificial and practical domains.

Understanding sequential events is important, for recognizing a sound sequence, diagnosing a system's varying fault-states, etc. For this purpose, the underlying coherence among observed events should be grasped. However, in many cases the variation speed of events is totally unknown and unpredictable. Therefore, previously presented and prevalent inference approaches using transition probabilities cannot be employed, and also there has been no former logical inference frameworks which overcome this crucial problem. This paper presents a Cost-based Cooperation of Multiple Abducers (CCMA), for explaining sequential events reflecting underlying common causes. Here, multiple abducers, i.e., inference systems of cost-based abduction each of which is assigned to one event observed at a time, work distributedly sending messages about their obtained hypotheses-sets to adjacent abducers. This CCMA obtains underlying common causes of sequential events, even if it is unknown and unpredictable how fast the underlying situation is varying, due to the mixture of numerical messages and logical inference of abduction without parameters on continuity of events like a conditional probabilities. We estimate the performance of this CCMA, for example problems of diagnosing varying faults of electronic circuits.

We analyze a learning method of weight of attributes in a similarity function for case retrieval by using relative distance information from a user. The relative distance information represents whether a training case is more similar to one case in the case base than to another case in the case base. We give an analysis in a PAC (probably approximately correct)-learning for the method. By using the method, we can efficiently learn weight such that the probability that the error rate of similar case retrieval by using the learned weight is more than ε is at most δ. The sample size of training cases to achieve the above is polynomially bounded in the number of attributes n, the size of case base, ε^<-l> and δ^<-l>, and the running time is polynomially bounded in the size of training cases. We also show experimental results on the sample size and the error rate for similar case retrieval under the assumption of uniform probability distribution over cases. The results indicate that the sample size is approximately 2n/ε on average.

Logic programs provide artificial intelligence (AI) with one of the most successful knowledge representations. Many applications in AI, however, need more expressive power and inferential efficiency than conventional logic programs. Therefore, many extended knowledge representation systems have been proposed. Intuitively, most of these are based on declarative paradigms and their theoretical foundations are similar. The declarative programs in this paper represent a class of programs which have been mathematically formulated with the intention of providing a unified theoretical foundation of such intuitively declarative knowledge representation systems. In the domain of logic programs, resolution is one of the most important inference rules, and the theory is well established. However, there is no established theory of inference rules for a more general class of declarative programs. In this paper three important inference rules for declarative programs, rules of specialization, cancellation and resolution, are discussed. The inference rule of specialization for declarative programs is almost the same as that for conventional logic programs, except for the applicability check of specialization. The inference rule of cancellation for declarative programs, on the other hand, is different from the conventional rule in that it is not always sound for all pairs of input clauses of declarative programs. The two inference rules, rules of specialization and cancellation, are combined to form the inference rule of resolution. New definitions of unifiers and safe unifiers are proposed to establish the inference rule of resolution for declarative programs.

This paper reports on a systematic construction method of an expert system for automatic software design. The systematic construction requires: 1. systematic knowledge acquisition; and, 2. systematic method for the design and the implementation. For the former, a software development organization with high maturity is taken as a target expert, and its hierarchical design work process is chosen as the knowledge model. The hierarchical design work process intersects with hierarchically organized documents. As design knowledge is thus hierarchically sectioned, systematic acquisition of the design knowledge becomes possible, resulting in a hierarchical framework with respective design process knowledge in each part. Based on this knowledge model, the model of an expert system for automatic design may be derived. Architecture design on this model is the first part of the systematic construction. For the latter, detail of systematic construction of an expert system unit, corresponding to a work process, is explained. Throughout the design, major Software Engineering practices are followed and further care has been taken to detail the design hierarchically in a small step. This systematic construction leads to a standardized procedures for actual design work. The expert system's characteristics of detailing of design and the learning effect of design knowledge are also reported.

An empirical study on referent identification requests or referring actions in multi-modal dialogues is reported. Through the examination of a corpus obtained by experiments in which experts explained the installation of a telephone with answering machine features in two situations, spoken-mode dialogue and multi-modal dialogue, referring actions in multi-modal dialogues were well analyzed and compared with those in spoken-mode dialogues from two perspectives: information communicated and the style of goal achievement. One of the findings obtained through this study is that the availability of pointing actions in multi-modal dialogues reduces the amount of information verbally communicated. The kind of information reduced from the spoken-mode situation depends on contextual status. This fact suggests that pointing actions are different from actions that communicate a specific type of information such as object location and figures. Rather, they are actions to activate the object referred to in the interlocutor's mental space. As communicating information on an object plays the same role in referring actions, pointing actions and verbal communication can be compared in this standpoint. Another finding is that in spoken-mode dialogues, as compared to multi-modal dialogues, the speakers realize identification requests as a series of fine-grained steps, and try to achieve them step by step with frequent confirmation.

KICK-SHOTGAN is known as the fast hypothetical reasoning system which avoids inefficient backtracking by forming a compiled inference-path network followed by the forward synthesis of necessary hypothesis combination along this network. In this paper, we propose an improvement of the KICK-SHOTGAN's subsumption process on the synthesis of some hypotheses using inference-path network.