We describe a new learning method, CLiP (Concept Learning from Typical Patterns), that performs induction over colored directed graphs. CLiP is capable of performing both inductive and deductive learning by mapping the problems into a colored digraph representation. In contrast to earlier approaches, CLiP uses a single learning algorithm to solve both kinds of problems. The learning procedure can be characterized as a variation of beam search guided by a simple, but effective, heuristic:typical pattern heuristic. We demonstrate the applicability of CLiP to the tasks of (1) inductive learning for classification and (2) deductive learning for efficient problem-solving. We show that the performance of CLiP on these tasks is comparable to that of standard approaches. Our preliminary results suggest that the generality of CLiP can be attributed to the expressiveness of the colored digraph representation which allows a number of seemingly different learning problems to be solved by a single algorithm. The other functions of CLiP and the limitations are also discussed together with the related work.

In this paper, we introduce a new accompaniment system which takes independence from the soloist into consideration. In order to make a natural-sounding ensemble, we allow our system to change its independence dynamically from the soloist according to the musical situation. We introduce three new techniques into our accompaniment system. "The performance plan" allows the system to perform with expression as a human player would. "The independence rate" corresponds to how independently the system should perform from the soloist. "The time of the ensemble" provides information to enable the system to correct its tempo according to the independence rate. In order to evaluate the system, we carried out a listening experiment where the subjects listened to three versions of the same musical piece. The first was accompanied by a real human being, while the other two were computer-accompanied, where the first system simulated past researches which lacked the techniques mentioned above, while the second incorporated them. The results of the experiment show that our system which incorporated the three techniques sounds better than the system based upon past researches, and that it even sounds as good as a human accompanist.

DCSP (Distributed Constraint Satisfaction Problem) provides a formal framework for studying cooperative distributed problem solving. Several algorithms for DCSP have been proposed. This paper presents a new method, LMO (Local Minimum driven Organizing), to solve DCSP. The basic algorithm is iterative improvement. Each agent measures the cost of its own instantiations as the number of violated constraints. If local change of instantiations reduces the cost, the agent performs hill-climbing locally. The advantage of this method is that agents solve their local problems in Parallel. 0ne drawback, however, is the possibility of getting caught in local minima. LMO provides a technique for escaping from local minima. It is summarized as follows. When an agent (A1) gets caught in a local minimum, 1) A1 sends its local CSP to the agent (A2) that shares a violated constraint with it, 2) A2 puts their CSPs together and solve them by a brute-force search.It eliminates local minima from search spaces, alters a hill-climbing for a brute-force search, and reduces communication overhead at the cost of the parallel execution. It is triggered by a local minimum, therefore the more local minima agents get caught, the more CSPs are put together and solved with a brute-force search. This means that the algorithm is adaptable to the problem it solves. In this paper, we verify this fact experimentally.

Minimal conflict search is applied to job shop scheduling. Job shop scheduling is not only NP-hard like many other combinatorial optimization problems, but is one of the hardest among them. Minimal conflict search, which is an efficient local search, prohibits transitions leading to solutions which include one of the previously generated critical paths. This mechanism allows the search to escape from local optimal solutions. In experiments on 10-job 10-machine problems, minimal conflict search was superior to simulated annealing which is one of the most powerful methods for job shop scheduling. In particular, minimal conflict search found the optimal solution of the notorious Fisher and Thompson problem. Unlike simulated annealing, minimal conflict search tends to find good solutions in the beginning of its execution, making the new algorithm of high practical importance.

In this paper, a new approach to solve an autonomous vehicle path planning problem is presented. Many of the current approach in this area involve creation of an explicit world model and effective search of optimal path in this model. One such world model decomposes space into a unique or hierarchical grid, and a classical search algorithm such as A searches the optimum path in the world model. Although this world decomposition provides efficient representation of space, it takes exponential computational cost corresponding to the size of space to search the optimal one. Addition to that, paths generated by planning systems operating within such representations tend to suffer from stair-stepping effects. Stair-stepping effect is a result of loss of space continuity resulting from decomposition of space into a grid, and it increases the complexity of vehicle guidance. This paper presents a path planning algorithm which can generate smooth paths that eliminates stair-stepping effect within a real-time. The algorithm is based on genetic algorithm, and can apply to various planning problems that require real-time planning capability in a dynamic domain.

Inductive inference is one of the most active work fields of machine learning. Various methods of concepts learning from examples are proposed. The method of "Version Space" proposed by T. M. Mitchell is a typical method of the concept learning from training examples. The algorithm of inference mechanism in Version Space, called candidate-elimination algorithm, needs a complete inference process to obtain the correct solution. Therefore, the algorithm has some problems, such as(l) it is very difficult for the algorithm to acquire disjunctive hypotheses, (2) the algorithm cannot form "plausible" hypotheses during the inference, (3) the algorithm cannot form hypotheses, if training examples have some wrong one, called noise, (4) the algorithm cannot form hypotheses from only positive examples, and (5) the algorithm cannot form hypotheses, if representation language has changed during the inference. We pay attention on the combinatorial structure of concepts form, then apply to Genetic Algorithm originally proposed by J. Holland [Holland 75]. In this paper, the method of the concept learning based on Genetic Algorithm is proposed. The new method can acquire disjunctive concepts and form plausible concepts during the inference. Concepts acquired by new algorithm have few influences of the noisy training data. As for the new algorithm, it was experimentally confirmed that the concepts acquisition only from positive examples is possible.

It has been accepted that the techniques of hierarchical problem solving can improve the efficiency of problem solving and theorem proving system. One of important problems in hierarchical problem solving is the automatic formation of the abstraction hierarchies. Some approaches have been proposed to form automatically the abstraction hierarchies for planning system. However, these approaches can not be applied to theorem proving system, since they are built on the planning systems based on the representation of STRIPS-like operators. In this paper we propose an approach to automatic formation of the abstraction hierarchies of theorem proving based on a definitional hierarchy. In this approach, each abstraction hierarchical level is assigned to a value based on the definitional hierarchy and the theorem to prove. A predicate will be supposed to be provable if the rank of the predicate in the definitional hierarchy is smaller than the value of the abstraction hierarchy. It is useful for the domains with a great amount of definitions such as mathematics. Contradiction may occur in the abstract theory generated by abstraction. To solve this problem a procedure of generating a consistent abstract theory by using propositional abstraction is proposed. This approach is implemented on a knowledge based theorem proving system in which axioms, definitions and theorems are represented as production rules. Some results of the experiment on topological space are investigated.

A hypothetical reasoning is an important knowledge system's framework because of its theoreticalbasis and its usefulness for practical problems including diagnosis, design, etc. One crucial problem with the hypothetical reasoning is, however, its slow inference speed. To achieve practical or tractable speed, an approximate solution method of O-1 integer prograrmming has been appliced so far to a weight-based hypothetical reasoning, where a numerical weight is assigned to each possible element hypothesis and the optimal solution hypothesis-set with the minimal sum of its element hypotheses' weights is searched. In order to take advantage of considering the knowledge structure of a given problem, the authors have previously proposed a network-based inference method called networked bubble propagation (NBP) method, which allows faster inference speed than that of the original approximate solution method of O-1 integer programming. In this paper, we present an improved NBP (I-NBP) method which has the preferreble property that it can always find the optimal solution (not near-optimal one) in polynomial-time if the knowledge structure is fairly simple including the case of singly-connected network. This new method can also achieve faster inference speed than the previous NBP method for finding a near-optimal solution when the knowledge stucture is in the general form. A theoretical study on the inference complexity related to the knowledge structure is also shown, and a new sufficient condition of polynomial-time inference is suggested.

This paper proposes a bottleneck improvement method on a new type of Queueing Network with Shared Stores (QNSS) using qualitative reasoning. Based on heuristics and knowledges obtained from evaluation experts, the authors have developed a new "qualitative reasoning" -based expert system, named BDES-SS (Bottleneck Diagnosis Expert System for QNSS), which can identify the bottlenecks, analyze the sources, and provide qualitative improvement plan option for QNSS. QNSS has the different types of bottlenecks in comparison with the ordinary QN. Primary types of bottlenecks in QNSS are "bottleneck Servers" with high utilization rates, "shared stores in overflow bottleneks", and "shared stores in underflow bottlenecks". In case of bottleneck improvement of QNSS with more than one bottlenecks, improvement Plans cause inconsistency one another. In order to improve all bottlenecks, BDES-SS first enumerates improvement plans for each of bottlenecks. Then, it selects a feasible Combination of the plans without inconsistency between bottlenecks. For this combination, the authors propose an approach by qualitative influence Propagation on the skeleton network which is a reduced from of given QNSS.

The ID3 algorithms provide robust inductive processes of learning concepts from examples by constructing decision trees. The standard ID3 algorithm, however, is restricted to utilize symbolic/numeric attributes, and receive non-structured examples. We here extend the algorithm so that it can treat hierarchical attributes, and can receive structured examples (A hierarchical attribute relates its values hierarchically, and a structured example is an example having more than one component). The first Problem is solved by finding adaptively appropriate values for getting a target decision tree based on the formulated value generalization process. The second is by introducing a new type of attribute-based descriptions in which any attribute refers to some specified components Computational experiments are also examined to show the validity of the proposed methods.