The target in computer has been changing from the numerical forms to the symbolic forms, which we human-being are more familiar with. Recently A. I. is studied as the central technique of computer science. In the near future Artificial Sense is thought to be studied as what follows A. I. This paper describes "Listening to the music", as an approach to Artificial Sense. When we have an end to simulate human-like process, signal processing and knowledge processing have to be concatenated. In this paper, the process "Listening to the music" is considered from three major stages ; transcription, analyzing music and understanding music. In the stage of transcription, the notes are extracted as symbol from acoustic signal. In the stage of analyzing music, the structures as melody, rhythm, chord progression are analyzed from the extracted symbol of the notes. In the stage of understanding music, the sentiments are extracted by the rules which describe the relation between the sentiment and the structure extracted in analyzing stage. This paper presents how machine gets the sentiments listening to the music performance.

This paper describes an expert system called SOCKS (Software Common Knowledge Transfer System). SOCKS is designed to transfer programming knowhow from skilled programmers to a knowledge base, and then to transfer the knowhow from the knowledge base to unskilled programmers. SOCKS stores 161 instructions, each specifying a piece of programming knowhow. Experience indicates that if an unskilled programmer's problem could be related to an instruction, it is easy for him to learn the instruction. Therefore, SOCKS allows a novice programmer to retrieve relevant instructions simply by choosing certain keywords of a problem. To construct the reasoning model of SOCKS, we assume the following concerning the skilled programmer who can consult which instructions are related to a specific problem : (1) He can associate concrete keywords derived from a problem with more abstract keywords ; (2) He knows the degree of strength of the relationship between each instruction and the keyword assigned to it. Based on this analysis, we can represent the following types of knowledge : (a) association rules which specify relationships between keywords, (b) relation rules which specify relationships between keywords and instructions, (c) keyword frames, and (d) instruction frames. The acquisition of such knowledge occurs in three stages : (i) initializing knowledge base is shared by skilled programmers. (ii) Comparing the solution of the initialized knowledge base with the skilled programmer's solution allows the knowledge base to be refined. (3) The tools are provided for the skilled programmers to maintain the knowledge base. Comparing SOCKS with the traditional keyword retrieval clarifies the fact that the association rules broaden the object for retrieval and prevent users from missing solutions and that relation rules make it easy to locate a solution among many output instructions. SOCKS is used for programming design, trouble-shooting, and training novice engineers with the aim of improving software products.

The progress of VLSI technology has made it possible to combine problem-decomposition methods and the parallel processing technology that runs concurrently a large number of processors. If a problem is decomposed into intermediate problems and they are all solved with processors, concurrently, then the efficiency of problem-solving will be improved remarkably. However, the traditional problem-decomposition methods are not fit for the parallel processing, because the initial state of each intermediate problems is determined by the solution of the one that precedes. In this paper, we propose a serial decomposition method and extend it to the one that is fit for parallel processing. In general, a serial decomposition method decomposes a given problem into the sequence of intermediate problems by setting subgoals. For this method to be of interest, it is necessary that the intermediate problems be solvable and be simpler than original one. To simplify them, we introduce equivalence relations on their state spaces and get quotient problems. We show a condition for the composition of solutions of these quotient problems to be a solution of the original problem. Furthermore we extend the method to the one that is fit for parallel processing. Since the initial state of an intermediate problem is contained in the goal of the preceding intermediate problem, if one regards all elements of the goal as candidates for the initial state and define intermediate problems for the candidates, respectively, the set of all intermediate problems surely contains the desired one. As a result, the number of intermediate problems to be solved is increased. The total processing time, however, will be improved, since they can be solved concurrently.

Natural language has many ambiguities, so the consideration of ambiguities is important for learning syntactic construction and semantics. In order to acquire the grammar and the meaning of words, it is necessary to have the flexibility and to gain the experience of learning by introducing not truth values (0 and 1) but a degree of uncertainty. So, for the purpose of modeling such a learning process, we introduced Dempster-Shafer's theory into our inductive learning system of natural language grammar. Our system, using CFG (Context-Free Grammar) as the syntactic representation and the semantic network as the semantic representation, can learn the grammar and the meaning of words with the epistemic (subjective) probability inductively, initially without presenting the grammar and the meaning of words.

We have developed PiL system:a learning system that can acquire strategy knowledge from training examples with EBG (Explanation-Based Generalization) method. EBG, however, is not sufficient in generalization, therefore many redundant training examples were needed for learning. The under-generalization results from the generalizatlon without SOLVABLE concept proposed by Mitchell et al., but the non-operationality of the concept prevents from its usage. Thus we propose how to define operational SOLVABLE concept and a new generalization method DSBG (Direct Solvability-Based Generalization), an EBG method extended with DS (Directly Solvable) concept that is an operational SOLVABLE concept. With DSBG, PiL generalizes the examples more sufficiently than EBG and the significant decrease in the amount of the training examples is possible.