Many of the studies on GAs give emphasis on finding the global optimal solution. In this paper, we propose a new method which extend the application of GAs to domains that require robust solutions. If a global optimal solution found is on a sharp-pointed location, there may be cases where it is not good to use this solution. In nature, the phenotypic feature of an organism is determined from the genotypic code of genes in the chromosome. During this process, there may be some perturbations. Let X be the phenotypic parameter vector, f(X) a fitness function and Δ a noise vector. As can be easily understood from the analogy of nature, actual fitness function should be of the form f(X+Δ). We use this analogy for the present work. Simulation results confirm the utility of this approach in finding robust solutions.

To acquire a real robot behaviors, it is important to learn in the real environment. Although most reinforcement learning researches have been made learning by simulation since the real environment learning takes large computation costs and also needs a lot of time. To realize the reinforcement learning of a real robot, it requires both an adaptation for the diversity of situations and a high speed learning method that can even learn from less trials. This paper describes the Realtime Reinforcement Learning for a Real Robot based on the exploitation oriented reinforcement learning method that it's learning cost is too small and has strictly incrementality to realize Realtime Reinforcement Learning with Automated Sub-Rewards Generation method by the abstraction of the task state to accelerate the learning process. Then the successive learning experiment in the real environment for the ball pushing task for the real robot is performed.

Browsing is one of the most popular ways to gather information in hypertext. The goal of the browsing task is often unclear, and the user's interest may change easily while the browsing. Therefore, techniques of information retrieval cannot be used directly to support the browsing task. However, this type of information gathering is gaining importance as digital libraries or WWW become more widespread. Apart from this, from the educational standpoint, several investigations have reported the significance of information gathering consistent with a learner's interests. In this paper, we propose a method to support the browsing task in hypertext where a user doesn't have clear goal and his/her interest may change easily while the browsing. The method infers the user's interest based on the user's browsing history. The browsing history reflects the user's interest and the user's current interest is formed through the browsing history. In this method, each node should have a set of keywords which describes the content of the node. We call the keywords indexes. Then, the history is described as appearance states of each index. The method calculates current weight of each index based on the appearance states. The weight means the user's interest for the index. Then, the weight of each node is calculated as the sum total of indexes included in the node. This method judges that the heaviest node is the most suitable node for the user's current interest. We have developed a browsing support system with this method for an encyclopedia provided as a CD-ROM. The result of an experimental evaluation of the system is also reported.

When Genetic Algorithms (GAs) are applied to optimization problems, characteristic preserving in designing coding/crossover and diversity maintaining in designing generation alternation are important. Generation alternation models are independent of problems, while coding/crossover depends on problems. We discuss generation alternation models in this paper. Simple GA is one of the well-known generation alternation models, however it has two problems. One is early convergence in the first stage of search and the other is evolutionary stagnation in the last stage of it. Many improvements and new models have been presented to overcome the above problems. In this paper, we propose a new generation alternation model called minimal generation gap (MGG) which has all advantages of conventional models. As generation alternation models use only information of fitness, alternation of generations can be regarded as a transformation of fitness distributions. We propose a new method of assessing generation alternation models. We measure the ability of avoiding the early convergence and suppressing the evolutionary stagnation by the dynamics of the best value and variance of fitness distributions. From the results of some experiments, we found that MGG is the most desirable model which can avoid the early convergence and suppress the evolutionary stagnation. We also show the efficiency of MGG by applying it to benchmarks in different two domains: function optimization and traveling salesman problems. In the both domains, MGG showed higher performance than the other conventional models especially under small population size.

This paper is concerned with techniques for analyzing binary data by iterations of weighted least squares method from the standpoint of comparison with hidden-layer feedforward neural networks. By use of arc-sine (angular) transformation for binary data with hidden-layer feedforward neural networks, we particularly discuss learning algorithms of regression models which relate qualitative variables to other variables through a logistic distribution functional form. Akaike's information criterion (AIC) provides us the decision as to which several competing network architectures for binary regression models. We contrast the merits of hidden-layer feedforward neural networks with those of ordinary logistic regression models.

The stable model semantics is now one of the standard semantics for general logic programs. A simple procedure for computing stable models has been proposed by Inoue, et al. However, the procedure may require exponential search space in the worst case. In this paper, we propose search strategies based on dynamic analysis, which is incorporated into the computation of stable models based on MGTP (Model Generation Theorem Prover). We also show some experimental results which show remarkable speedup compared with the original procedure.

In human-computer dialogue, a diverse range of situations arise from uncertainty in reasoning regarding the user's beliefs and plans. The dialogue system is required to plan the contents of its utterances under such uncertain conditions. This paper presents a decision-theoretic framework of content planning for dialogue systems. In our framework, the uncertain domain knowledge and user's model are both declaratively described as probabilistic constraints, which represent uncertain situations as probabilistic distributions over possible worlds. We estimate this probabilistic distribution using a Bayesian network. Given a utility function to evaluate the degree of goal attainment in each possible world, the system chooses the content of its utterance based on the expected utility maximization principle. Our framework can perform content planning under uncertain conditions without any procedural planning strategy.

In this paper we propose a match algorithm for production systems, which is a method for dealing with expensive productions. Production systems are an established method for encoding knowledge in an expert system. Expensive productions are rules which would be required the extraordinary time and space to match (compare the left-hand side of each rule against the current working memory). This paper presents the improvement of a direct match algorithm for production systems, using one dimensional arrays stored an attribute value of a working memory element. It is a method for reducing computer loads which are caused by expensive productions.