One of the causes of inefficient problem solving for planning based on means-ends analysis is cancel of goal, which means the action that planner achieves a sub-goal and cancels it and achieves it again. In this paper, we propose the conflict resolution rule (CRR) as a heuristic for efficient problem solving, which directs to select an appropriate sub-goal in order to prevent the cancel of goals. We also propose a learning method of CRRs, in which they are obtained by analyzing sequences of operators which include the cancel of goals. Since CRRs are obtained without any domain-dependent knowledge and instructions of a teacher, they are able to be efficiently applied to larger scale or more complicated problems. We examined the availability of CRRs with "Robot Key Lock Problem". As a result of experiments, we showed to succeed in obtaining useful CRRs from a few number of training examples, and realizing efficient problem solving by generating efficient sequences of operators.

A machine learning technique called Graph-Based Induction (GBI) efficiently extracts typical patterns from directed graph data by stepwise pair expansion (pairwise chunking). In this paper, we apply GBI to the classification problem by interpreting the root node as a class node and the links attached to it as the primary attributes. The node at the other end of each link is the value of the attribute, which has secondary attributes. Thus, each attribute can have its own attributes recursively, and the graph becomes a directed tree. In this case, the pairwise chunking must start at the root node and go backwards following the links. Though inducing classification rules by GBI has already been proposed, what is new in this paper is that we have changed the evaluation criterion in selecting the pair to be chunked and also added the pruning mechanism to GBI in order to avoid overfitting to the training data. The new classification version of GBI, called CL-GBI, is applied to two kinds of problems. One is the command prediction problem where CL-GBI must predict what the user wants to do next, and the other is the WWW browsing problem where CL-GBI must extract chracteristic routes from the access log, i. e. the browsing history of many users. The results indicates the effectiveness of our approach.

The Baldwin effect is known as an interaction between learning and evolution, which suggests that individual lifetime learning can influence the course of evolution without the Lamarckian mechanism. Since Hinton and Nowlan showed this by computer simulation, many studies have been conducted in the static environment. Our concern is to consider the Baldwin effect in dynamic environment, especially when there is no explicit optimal solution through generations and it only depends on interactions among agents. We adopt the iterated Prisoner's Dilemma as a dynamic environment and introduce phenotypic plasticity to strategies by using a meta-learning rule termed "Meta-Pavlov". In this simulation, the Baldwin effect was observed as follows : First, strategies with enough plasticity spread, which caused a shift from defective population to cooperative. Second, these strategies were replaced by the strategy [x00x] which has a modest amount of plasticity. We have analyzed this strategy and have shown that it satisfies the ESS (Evolutionarily Stable Strategy) condition and establishes robust and cooperative relationships in the population.

In conventional studies, metonymy interpretation has been carried out by using a hand-built database that includes relationships between words concerned with metonymy, such as a special knowledge base of metonymy and a semantic network. However, these relationships between words are diverse, and it is difficult to manually make a detailed database. Therefore, in this paper we interpret metonymy by using examples in the form of noun phrases such as "Noun X no Noun Y (Noun Y of Noun X)" and "Noun X Noun Y, " instead of a hand-built database. This method has two advantages. One is that a hand-built database of metonymy is not necessary because we use examples. The second is that we can interpret newly-coined metonymies by using a new corpus. In experiments using this method on 23 metonymy sentences taken from textbooks, we correctly judged 17 sentences to be metonymy sentences and correctly interpreted 7 of them.

The purpose of this paper is to propose an algorithm that computes all the minimal models in propositional prioritized circumscription by 0-1 integer linear programming. An algorithm for non-prioritized circumscription based on integer linear programming has been proposed by Bell et al. We extend their method to handle prioritized circumscription. A set of formulas is transformed into a set of linear inequality constraints and the objective function is determined by priority over minimized propositions. Our algorithm works as follows. We firstly compute a minimal model by optimizing the objective function under the given constraints. Then, we add a set of new constraints which prevents from choosing models that is not smaller than the previous minimal model and we iterate this process until any models are derived from the constraints.

Although many diagnostic expert systems have been developed, only a few of them became of practical use mainly because of "knowledge acquisition bottleneck". We developed a plant diagnosis system using model based reasoning (MBR). The system contains a model of sequence control logic which can be automatically derived from design data, thus resolving the knowledge acquisition bottleneck. A plant has a huge number of sensors which are prone to failure. The system starts with abnormal observed values, traces through the sequence control model backward, and identifies failed sensors. The reverse tracing involves combinatorial explosion, especially the data is time series as in our system. In order to make the system practical, we proposed the following methods : pruning branches based on observation values, focusing a time interval of time series data, and selecting a physically plausible solution among logical candidates. Experimental results showed that the system could detect accurate solutions within reasonable time.

There are two types of weight assignment in COP (Constraint Optimization Problem), namely using node weight and arc-constraint weight. Most COP methods developed so far deal with one type of weight, and cannot treat another one. In this paper, we propose a weight conversion method between the two types. With this conversion, a COP method developed for one type can be applyed successfully to another type. This is very advantageous to develop efficient COP methods.