Classical mathematical logic has a well-known problem of "material-implicational paradoxes", and therefore it is not suitable as the logical basis of knowledge engineering. Relevant logics are constructed in order to find a mathematically satisfactory way of grasping the notion of entailment. They are free of material-implicational paradoxes and it seemed that they are suitable as the logical basis of knowledge engineering. However, recently it is pointed out that there still are other types of paradoxes, called "conjunction-implicational paradoxes"and "disjunction-implicational paradoxes"in relevant logics. In order to adopt relevant logics as the fundamental theory of knowledge representation and reasoning, it is necessary to exclude the conjunction-implicational and disjunction-implicational paradoxes from relevant logics. In this paper, we investigate the formal characteristics of conjunction-implicational paradoxes and disjunction-implicational paradoxes, and propose a necessary condition, named the strong relevance principle, for paradox-free relevant logic systems which are free of not only material-implicational paradoxes but also conjunction-implicational and disjunction-implicational paradoxes. We show that relevant logic systems Rc, Ec, and Tc, which are proposed in order to construct a paradox-free relevant logic system, satisfy the strong relevance principle.

A new theory of representation and computation for multiset domains was developed. A multiset (or a bag) is a generalization of sets, which allows multiple occurrences of an element. This multiset is useful in representing the collection of objects, and it can be often used as a convenient alternative for sets. In this paper, we investigate the simplest case of a multiset, i.e., the elements in the multiset discussed here are symbols with no structure. While conventional theories for multisets are based on the logic paradigm, this new theory is based on the equivalent transformation(ET)paradigm, where computation is regarded as "equivalent transformation of declarative descriptions." The most important difference from the logic paradigm is the existence of a strict and general foundation. In the ET paradigm, a class of declarative descriptions, called "constraint declarative programs"on "specialization systems, "is used for all possible domains. A generic definition of declarative semantics for all constraint declarative programs is established. A specialization system is a mathematical structure that characterizes each domain. A class of specialization systems is proposed to formalize multiset domains. To delay unnecessary unification, a new transformation rule, called expanding, and equivalent transformation rules for equality constraints are used instead of unfolding. Basic propositions justifying equivalent transformation rules for equality constraints on multiset domains are proven.

Most of the reinforcement learning (RL) algorithms assume that the learning processes of embedded agents can be formulated as Markov Decision Processes (MDPs). However, the assumption is not valid for many realistic problems. Therefore, research on RL techniques for non-Markovian environments is gaining more attention recently. We have developed a Bayesian approach to RL in non-Markovian environments, in which the environment is modeled as a history tree model, a stochastic model with variable memory length. In our approach, given a class of history trees, the agent explores the environment and learns the maximum a posteriori (MAP) model on the basis of Bayesian Statistics. The optimal policy can be computed by Dynamic Programming, after the agent has learned the environment model. Unlike many other model learning techniques, our approach does not suffer from the problems of noise and overfitting, thanks to the Bayesian framework. We have analyzed the asymptotic behavior of the proposed algorithm and have proved that if the given class contains the exact model of the environment, the model learned by our algorithm converges to it. We also present the results of our experiments in two non-Markovian environments.

Cost-based abduction is a well established formalization of abduction for obtaining the most desirable hypotheses-set. However, slow inference-speed is its crucial problem-it is NP-complete. In order to overcome this problem, mathematical programming based methods have been shown to be effective. However, their performances were mainly evaluated empirically, not theoretically supported. Recently, a new condition was presented under which polynomial time cost-based abduction can be achieved by solving the relaxed linear programming problem, though inconsistencies among hypotheses were not considered. This paper presents a new sufficient condition for the polynomial solvability of cost-based abduction by linear programming relaxation, considering inconsistencies among hypotheses which heavily influence the time complexity of abduction. Testing the new condition is possible within polynomial time. Our theory is based on the constraint network of a previously presented Networked Bubble Propagation (NBP) method, a high-speed approximate cost-based abduction method.

Engineers are often confronted with the problem of extracting information about poorly-known processes from data. Discerning the significant patterns in data, as a first step to process understanding, can be greatly facilitated by reducing dimensionality. An artificial neural network can develop a compact representation of the input data. This neural network can be applied discerning the significant patterns in data. The neural network, which can reduce dimensionality and produce a feature space map, contains an internal "bottleneck"layer (containing fewer units than input or output layers) and two additional hidden layers. In the case of this type neural network, the inputs of the network are reproduced at the output layer. An important problem is to determine the optimal neural network architecture to acquire the optimal feature space map. This paper proposes that information criteria are applied to determine the optimal neural network architecture and presents that Neural Network Information Criterion (NNIC) is the most useful for acquisition of the feature space map with a simple simulation. The neural network, which is selected by using NNIC, can adequately reduce dimensionality and produce the optimal feature space map resembling the actual distribution of the underlying system parameters.

Recently, the software has become very large scale and in addition its structure has become very complex. Hence, it is required for software managers and users to have considerable knowledge concerning the software systems and furthermore to perform many manual operations for maintaining and controlling them. In this paper, based on the notion of policy, we consider a sophisticated way to resolve such a serious problem. A policy description language is proposed to support the automatic software operations in which a policy is described in an abstract level using IS-A hierarchy of knowledge. Furthermore, we present the design to implement the policy system, and discuss its application to the Inter-Domain Routing Protocol(IDRP).

Several techniques have been developed to solve puzzle problems in conventional AI, but there are few attempts to compose problems automatically by computers. Tsume-Shogi, a mating problem of Japanese Chess, is a kind of puzzles that is created and solved according to specific rules. This paper presents a system to compose Tsume-Shogi problems by reverse method. The search space increases enormously when the reverse method is adopted, but we can reduce it by using some constraints. We conducted several experiments with our method to compose Tsume-Shogi problems and showed that our system could compose some good short Tsume-Shogi problems and some special Kyuku-Tsume problems.

It is important to guide a user to interesting documents in a large-scale document-database. However, when the user is not an expert of the area of his/her new interest, it is difficult for the user to name precise keywords in which he/she is interested, nor to select areas of his/her own interest. This paper presents an Index Navigator which clarifies what areas the user is interested in, what keywords he/she should investigate, and what documents concern his/her interest. A tough problem for such a system is to understand interesting areas for the user, among other areas-sets which can explain his/her behaviors. Our Index Navigator employs an inference method called Cost-based Cooperation of Multiple Abducers (CCMA), for understanding user's interest from the history of the user's expression of interest in insufficient keywords, even if the changing speed of the user's interest is totally unknown. With this device, the Index Navigator guides the user to really important areas, keywords and documents.

This paper discusses analytically and experimentally an optimal structure of a large scale knowledge base of words, which is automatically constructed from machine-readable dictionaries. In this knowledge base, each word is represented by a series of weighted keywords. The keywords have some relationship with the word, and the weights of the keywords represent the degree of the strength of the relationship between the word and keywords. In constructing this kind of knowledge base, it is important to select the optimal set of keywords used to represent every word in the knowledge base, considering the ability of measuring the semantic similarity between words. Our analysis, using a simplified model of the knowledge base based on probability theory, has shown that a smaller keyword set using the higher level keyword in the conceptual hierarchy becomes optimal when the size of the knowledge base, namely, the total number of words in it or the average number of keywords per word, becomes large. On the other hand, an experiment using six knowledge bases modified from the previously constructed knowledge base of 40000 Japanese daily-used words has verified the existence of the optimal keyword set. This means that the above mentioned analysis is useful in the design of a knowledge base in which each word is generally represented by a vector. In addition, we have found, from both a subjective evaluation based on human judgment and a newly proposed objective evaluation using a published synonym dictionary, that a set of about 2000 keywords is optimal for constructing a knowledge base of this size.

This paper describes a sentence alignment environment, BACCS (Bilingual Aligned Corpus Construction System). BACCS incorporates our robust alignment system which is powerful and effective for translations between Japanese and English. Moreover, BACCS offers a graphical user interface that enables a corpus-builder to easily confirm and modify alignment results through our novel alignment system. This interface also offers the function of easily extending the system dictionary. Experimental results confirm that users can align bilingual corpus by sentence-level with BACCS about twice faster than without it.