CAFE is a system which achieves concept formation incrementally while considering relations between objects of complicated structure. Objects in a domain of CAFE are represented not only by simple attribute-value pairs (as in previous systems) but by relations to another objects. In our system, Incremental hill-crimbing method is used to create a concept hierarchy. To improve this method, we propose a new mechanism called mutual induction. This mechanism enables the system to create a concept hierarchy which appropriately reflects the information of relations between objects. We have tested CAFE in a simple artificial domain. Our experiments have clearly proved the effect of mutual induction.

The inductive decision tree learning system can automatically develop some classification rules from a set of training examples, and has been applied in many application domains. This paper presents an approach, named INDECTS, for dealing with continuous attributes such as numerical data in the inductive decision tree learning. INDECTS is a variant of ID3 algorithm and has a labeling procedure for the numerical attributes. The labeling procedure divides all numerical data into several clusters, and makes a new discrete symbolic attribute that has several data thresholds for testing a numerical attribute. The procedure consists of a dividing operation and a unifying operation for data clusters. The key idea of this algorithm is to collect numerical data into data clusters according to its classification class, and to make thresholds for the adjacent data clusters by calculating maximum information gain. The labeling procedure is executed in each decision tree expanding step, so that each numerical attribute is dynamically translated into the discrete attributes, which have high information gain as the classification criteria. Some experiments show that this algorithm is more powerful and available for many types of numerical data than the existing methods.

We propose an instance-based learning algorithm named IBRL3 which learns how to avoid the negative reinforcement from environment. A cart-pole balancing problem and a monitoring ship navigation problem are used to certify its learing performance. In this algorithm, a tuple of input and output data of each execution cycle are stored in memory verbatim, and the action of each cycle is decided by retrieving the nearest neighbor of the current input data. The number of stored instances is reduced by replacing the nearest but less reliable instance by new one. Experimental results of computer simulation show that IBRL3 is robust for distinct settings of parameter and for noisy environments, and it is efficient enough to apply to real-time control problems.

Search is one of the most important and basical technique underlying most computer problem solving methods. Exhaustive search method explore all possible paths to a goal state. However, this method requires exponentially large time and space for computation as state spaces become large. Problem decomposition is one of the effective methods to reduce search space. General Problem Solver implements means-ends analysis. The necessary condition for its applicability is the existence of a set of subgoals and an ordering among them, such that once satisfying a subgoal, it must not be violated in order to satisfy the remaining subgoals. But problems, like Rubik's Cube, whose subgoals have strong interaction among one another, do not satisfy this condition. Korf has developed Macro Problem Solver and used macro operators to overcome this difficulty. His system solves a problem based on a macro table without no search. A macro table is a table of macro operators whose column headings are state components (differences in GPS) and whose row headings are the values of state components. Korf proposes a method to learn macro tables. His idea is basically a backward search from a goal by applying the inverse of the primitive operators. But in this paper, we consider the problem of acquiring a macro table from traces. We first define the extended definition of Korf's Macro Table, called Problem Decomposition Strategy. Because the acquisition of it seems to be intractable in general, we limit the attention to the domains which are information preserving and whose goal state space is composed of only one state. After investigating the characteristics of such domains, we present the efficient method for acquiring Problem Decomposition Strategy from traces. Our method does not need the information of subgoal ordering, which is necessary in advance for applying Korf's method. This is one of the most important contributions of this research.

Explanation-Based Learning (EBL) has been used for speed-up learning in problem solving. Since there ary many combinations of macros in each explanation, EBL systems need a selective learning mechanism of macros. Some systems select macros that connects two peaks in a heuristic function. Another system employs heuristics that select useful macros. Although they work well in some domains, such methods depend on domain-dependent heuristics that have to be exploited by their users. This paper presents a heuristic-independent mechanism by detecting backtracking. The method uses a dead-end path as a negative explanation tree, compares it with positive one, and finds a first different node to remove its corresponding rule by composing a macro. Repeated substructures in such a macro are then combined by applying the generalization-to-N technique and by sharing common substructures. Experimental results in STRIPS domain show that, by selecting an appropriate set of macros, (1) backtracking in solving training examples are suppressed, (2) its problem solving efficiency does not deteriorate even after learning a number of examples, (3) after learning 30 training examples, no backtracking occurs in solving 100 test examples different from the training examples. In conclusion, the proposed method speeds up the problem solving from 10 to 100 times.

Learning systems find a concept based on positive and negative examples by using given terms, such as features and predicates. Most learning systems employ selective induction, and find a concept description composed of only predefined terms. However, if such terms are not appropriate, constructive induction or shift of bias are required to invent new terms. Although there has been increasing interest in systems which induce a first-order logic program from examples, there are few systems that perform constructive induction. FOCL invents new terms, i.e., new predicates, by combining existing subpredicates. Based on interaction with its user, CIGOL invents terms without any given subpredicate. This paper presents Discrimination-Based Constructive inductive learning (DBC) which invents a new predicate without any given subpredicate nor any user interaction. Triggered by failure of selective induction, DBC introduces a new predicate into a previously found incomplete clause. This is performed by searching for a minimal relevant variable set forming a new predicate that discriminates positive examples from negative ones. If necessary, DBC also recursively invents subpredicates for the definition. Experimental results show that, without interactive guidance, our system CHAMP can construct meaningful predicates on predefined ones or from scratch. Our approach is system independent and applicable to other selective learning systems such as FOIL.

Knowledge acquisition is known as the "bottleneck" of expert system construction, because of its difficulties. Recently, the importance of knowledge acquisition has been recognized, thus it has become a focus of much research. Ideally, knowledge acquisition systems should acquire not only the experts' domain knowledge but also their heuristics. Furthermore, such expertise should be elicited with less burden of the expert. Interview is an effective technique of acquring knowledge. But systems using the technique produce questions that are lengthy, troublesome and difficult to answer for experts. Therefore, to reduce the redundancies of such questions, we propose a knowledge acquisition technique which integrates the method of learning from examples using EBL framework and the method of acquiring knowledge with interview. In this paper, we focus our attention on the above two acquiring methods and their integration. To realize the idea, we choose oil hydraulic system design as the task domain, because it is easy to understand and prepare its related domain knowledge. We have completed implementation of the learning module based on EBL. First, our basic idea is stated clearly. And overview of the system architecture is presented. Then behavior of the learning module is illustrated with some concrete examples of acquired design knowledge. Two interactions conducted by the interview module are also presented.

Genetic Algorithms (GA) is a new learning paradigm that models a natural evolution mechanism. The framework of GA straightly corresponds to an optimization problem. They are classified into functional optimization and combinatorial one, and have been studied in different manners. GA can be applied to both types of problems and moreover their combinations. According to generations, GA will discover and accumulate building blocks in the form of schemata, and find the global solution as their combinations. It is said GA can find the global solution rapidly if the population holds sufficient varieties. However, this expectation has not been confirmed rigidly. Indeed, there are some problems pointed out such as the early convergence problem in functional optimization, and the encode/decode-crossover problem in combinatorial one. In this paper, we give a solution to the encode/decode-crossover problem for traveling salesman problems (TSP) with a character-preserving GA. In section 2, we define the encode/decode-crossover problem. The encode-decode problem is to define a correspondence between GA space and problem space. The crossover problem is to define a crossover method in GA space. They are closely related to the performance of GA. We point out some problems with conventional approaches for TSP. We propose three criteria to define better encode/decode ; the completeness, soundness and non-redundancy. We also propose a criterion to define better crossover ; character-preservingness. In section 3, we propose a character-preserving GA. In TSP, good subtours are worth preserving for descendants. We propose a subtour exchange crossover, that will not break subtours as possible. We also propose a compress method to improve efficiency. In section 4, we design an experiment to confirm usefulness of our character-preserving GA. We use a double-circled TSP in which the same numbers of cities are placed on two concentrated circles. There are two kinds of local solutions ; "C"-type and "O"-type. The ratio between outer and inner radius determines which is the optimum solution. We vary the radius ratio and see how much optimal solutions are obtained. In the result, character-preserving GA finds optimal solutions effectively.

One of the main task yet to be done in discourse comprehension is to figure out the information which we could not know by syntactic and semantic analysis. Among them are the referents of pronouns and the structure of discourse. These are found by anaphora resolution and discourse parsing respectively. However the basic intuition in this work is that anaphora resolution and discourse parsing are not separate businesses but mutually dependent tasks in processing. The process of constructing discourse structure and determining the discourse content are not able to be sequentially solved in general. To solve this situation, we introduce a different kind of factor that gives us another kind of clue to solve the problem, which we call activity propagation in this paper. Roughly speaking, we assume every referent has its activity. On the other hand, in sentences there are expressions that are willing to connect to already existing referent. We call this kind of expression "hungry referring expression". Using these notions we formalize the anaphora resolution as a link establishing process between referent in high activity and hungry referring expressions. Hopefully all these notions and ideas would be conflated into one whole structure to solve the problem by constraint satisfaction paradigm.

An evaluation method for a Japanese to English machine translation system which estimates translation quality based on translation technologies and source text characteristics is proposed. This method is applied to the translation of newspaper articles by the ALT-J/E system and further problems that need to be surmounted. Assuming 70% acceptability to be a practical level of quality, it is shown that morphological and dependency analysis technologies would require accuracy rate per word and per "bunsetsu" (which are Japanese phrase, in most case, consisting of a noun and "joshi" or post-positional word) as high as 99.8% and 99.4%, respectively. Current technology has reached this level. But the precision of translation for individual expressions is lower than expected values and the acceptability rate for the entire sentence drops to between 40% and 50%. To achieve an acceptability rate of 70%, it is necessary to achieve an average 7% improvement in 9 types of expressions and thus achieve a successful translation ratio of 96%. To upgrade the translation rate of these expressions, it will be necessary to establish techniques for meaning analysis of post-positional words, translation of expressions that combine noun clauses and compounded words, handling connections and embedded clauses, and verification of ellipses and paratactical constructions.

In business expert systems for examination tasks, advice tasks and so on, not only rules but cases are necessary for decision-making. In a conventional way, rules are treated by the rule-based reasoning (RBR) and cases are treated by the case-based reasoning (CBR). However, both reasoning paradigms are separated and to use them together is difficult. In this paper, we propose an method for integrating both reasoning paradigms. The proposed method has following characteristics : (1) Knowledge representation and processing method of knowledge for CBR, similar with RBR. (2) Dynamic selection of the relevant reasoning paradigm according to the working memory (WM). (3) Reasoning method which starts form arbitary WM and produces plural solutions if there are any. Through devolopment and test of intelligent form-fillng front-end utilizing the proposed integrating method, it is demonstrated that knowledge including rules and usage of cases is defined easily, that a relevant paradigm is selected according to the WM, and that it supports users' decision-making by proposing alternatives.

In this paper we present a machine learning system Rhea. Rhea is characterized as a cross-domain concept formation system that obtains concepts from tuples of instances from multiple domains. Traditionally, concept formation is modeled in single domains. However, concepts should be formed with their use in mind. For example, the cancept of "chair" cannot be formed completely without the reasoner having a goal to sit on it. This task can be viewed as a cross-domain concept formation between form and function (goal). When learning from n domains, Rhea accepts n-tuple inputs. Each fragment of the input is an instance from one domain. Rhea has a conjecture that some rules constrain the possible combinations of instances from different domains and tries to find and/or generalize the rules. We implement Rhea in the domains of natural language expressions and outer-world descriptions. The implementation can acquire (1) extensions to the representation language for natural language expressions, which are syntactic rules that parse input expressions, (2) rules that give an account for relations between the domains, which can be interpreted as "meaning" of language expressions and classes of language expressions, and (3) classes each rule can apply, which are categorization of the input language expressions. To sum up, the implementation can be seen as a unified model of syntax and semantics acquisition based on general learning mechanism and shows learning from two domains is essential to classifications and finding rules.