We have implemented a simple model of puns in a program (BOKE) which generates puns in Japanese, using linguistic information from a general-purpose lexicon. Our rough evaluation indicates that the puns generated by the program are of comparable quality to those generated by humans. BOKE differs from an earlier English-language system (JAPE) only in the lexicon and the templates used to generate the surface text-the punning mechanisms are the same. This suggests that our model of puns is language independent.

=g. Secondly, it is proved that the set L(P, q) can be transformed into another form f(M(P')), i.e., it is represented by declarative semantics of a program P' determined by P and q. Finally, equivalent transformation (ET) is proposed as a new m

=qθ in terms of inference. The theory of SLD resolution provides Prolog with a theoretical foundation, proving that SLD resolution is a sound and complete procedure for logical problems. It should be pointed out, however, that SLD resolution is not the be

Equivalent transformation seems to be one of the most fundamental methods for problem solving since it strictly ensures the validity of a solution as long as the given problem is transformed equivalently in each step of computation. There exists, we believe, no other theory that can ensure the validity of problem solving both generally and straightforwardly. This paper presents a foundation of "rule-based equivalent transformation", where a given problem is solved by successive equivalent transformation in terms of independent and valid rules. A general theory of equivalent transformation is developed, where definitions of basic concepts such as representation systems, formalization of problems and equivalent transformation rules, are introduced. Based on this foundation, we also introduce a specific class of declarative descriptions called declarative programs. A declarative programs is similar to a logic program in the sense that it is a set of definite clauses consisting of atoms, except that the atom-substitution structure in logic is extended to an axiomatic structure. Since the class of declarative programs is a general extension of logic programs, equivalent transformation of declarative programs can be used to solve various problems. A string rewriting problem is used to explain how declarative programs can be used for formulating problems, where declarative programs on a string domain, not logic probrams on a usual term domain, provide the most natural representation for the problem. Experimental results of a natural language understanding system show that the efficiency of computation may be increased by using (1) equivalent transformation rules not based on unfolding, (2) data structures other than terms, and (3) flexible control of computation.

Hypothetical reasoning is an important framework for knowledge based systems due to its theoretical basis and its usefulness for many practical problems. Since its inference time grows exponentially with respect to problem size, its efficiency becomes the most crucial problem when applying it to practical problems. Some approximate solution methods have been proposed for computing cost-based hypothetical reasoning problems efficiently;however, their mechanisms are complex for human to understand. We here present an understandable efficient method called SL(slide-down and lift-up) method, which uses a linear programming technique, namely simplex method, for determining aninitial search point and a non-linear programming technique for efficiently finding a near-optimal 0-1 solution. To escape from trapping into local optima, we have developed a new local handler, which systematically fixes a variable to a locally consistent value when a local optimal point is detected. This SL method can find a near-optimal solution for cost-based hypothetical reasoning in polynomial time when respect to problem size. From pictorially illustrated behaviors of the SL method, its simple inference mechanism can be easily understood.

Sound source identification is an important but difficult problem in sound source separation. It is also a problem in the symbolization of music performances which include multiple simultaneous notes. As a solution to this problem, this paper presents a new method that can significantly improve the precision of sound source identification for music. Identification is here defined as the recognition of instrument names for each note included in an ensemble music monaural (or stereo) signal. The key idea of the proposed method is utilizing musical context. First we define the "music stream" that corresponds to a sequence of notes as a basic representation of musical context. We then describe the Bayesian method to introduce the contextual information to sound source identification. Experimental results show that the proposed method improves the accuracy of the source identification task for three-part ensemble music signals from an average of 67.8 % to 88.5 %.

Reinforcement learning (RL) is the class of learning to obtain a policy to interact with the environment among an autonomous agent, only with the clue of the signal which tells the agent whether former interactions were adequate or not. Most RL algorithms are directed to obtain an optimal controller, which specification is unreasonable and often in vain because of the contradiction between exploration and exploitation. This paper proposes a new framework of RL, satisficing RL, shows that directing to satisfice is a reasonable specification free from contradictions, and also presents an RL system, which is mathematically assured to satisfice only with next to the least constraints. An example presented will help us to grasp the idea of satisficing RL. The assurance of satisficing is described as a convergence theorem. Other features of the RL system are also described, while convergence rate estimation is reserved as a future work. Since we know the real world includes a great amount of states, in discussing the real problems we should assume the state set to be infinite. On the other hand, work memories are necessary for the agent to be intelligent, which are made to contain the information about the environment. For this reason, this paper also proposes the way to satisfice in the environment with perceptual aliasing, using finite memories efficiently.

The formalization of human deductive reasoning is a main issue in artificial intelligence. Although classical logic (CL) is one of the most useful ways for the formalization, the implication of CL has some fallacies. For example, in CL, A ⟶ B can be inferred from B for an arbitrary formula A. This inference is not relevant from the viewpoint of the meaning of implication which human has. In human reasoning, when A ⟶ B is inferred, A and B should be related. Relevant logic has been studied for removal of the implication fallacies of CL. For the relevance of A ⟶ B, several principles are introduced. One of the most important principles is Variable-sharing, where, if A ⟶ B is a theorem, then A and B share an atomic proposition. Relevant logical system should satisfy this principle. Another principle is that the truth values of A and B do not decide A ⟶ B. Classification of the fallacies of implication is also introduced. Fallacies are classified into those of relevance, validity, or necessity. Since the fallacies of relevance and validity are strong fallacies, they are removed from almost all relevant logical systems. Relevant logic, however, is a weaker logical system than CL. In this paper, we propose the relevant logic ER. Then we prove that Variable-sharing holds in ER, that fallacies of relevance and validity are removed from ER. We also prove that ER is not weaker than R. Respecially, disjunctive syllogism holds in ER but does not hold in R. In this sense, ER is a natural formalization of human reasoning.

One central problem in AI is to compare a structural situation with another situation. Comparison makes it clear which parts of the structures are different and/or considered to be the same. Numerous methods were proposed in relation to analogical reasoning, case-based reasoning, etc. Most of them, however, lack sufficiently formalized treatment of background knowledge defining the semantics in the structures. We discuss a general framework for structural comparison, and then formulate under this framework a method of comparing concept structures by introducing the notion of coverning. It should be noted that, in a concept structure, the concepts specified by background knowledge based descriptions are partially ordered by a generic/specific relation found in background knowledge. The comprison, called here BK-comparison, of two concept structures proceeds by finding a correspondence which (1) satisfies that if the concepts in one structure have a generic/specific relation, their corresponding concepts in another structure do have the same relation, and (2) produces an MSC-covering defined as a set of highly specific generalized descriptions, each of which refers to two corresponding similar concepts in the different structures. The BK-comprison is speeded up by a new beam-search method that employs the facility of the summed levels to estimate whether or not the found coverings are the MSC-coverings, and the directional schemata guiding the search toward a successful direction. The proposed method is verified computationally by comparing concept structures induced from artificial and real-world data. Its applicability to various subject areas is also discussed.

Many of the studies on GAs give emphasis on finding the global optimal solution. If a global optimal solution found is on a sharp-pointed location, there may be cases where it is dangerous to use this solution. In the previous paper, we have proposed the GA/RS^3(GA with a Robust Solution Seraching Scheme) which extend the application of GAs to domains that require robust solutions, and its analytical and experimental study has been made limitting the search space to be one dimensional case. In this paper, we extend it to multidimensional case.