Akira Notsu

FCM classifier for high-dimensional data

A fuzzy classifier based on the fuzzy c-means (FCM) clustering has shown a decisive generalization ability in classification. The FCM classifier uses covariance structures to represent flexible shapes of clusters. Despite its effectiveness, the intense computation of covariance matrices is an impediment for classifying a set of high-dimensional data. This paper proposes a way of directly handling high-dimensional data in the FCM clustering and classification. The proposed classifier without any preprocessing outperforms the k-nearest neighbor (k-NN) classifier with PCA on the benchmark set of COREL image collection.

Fuzzy c-means classifier with particle swarm optimization

Fuzzy c-means-based classifier derived from a generalized fuzzy c-means (FCM) partition and optimized by particle swarm optimization (PSO) is proposed. The procedure consists of two phases. The first phase is an unsupervised clustering, which is not initialized with random numbers, hence being deterministic. The second phase is a supervised classification. The parameters of membership functions and the location of cluster centers are optimized by the PSO and cross validation (CV) procedures. Since different types of classifiers work best for different types of data, our strategy is to parameterize the classifier and tailor it to individual data set. The FCM classifier outperforms well established methods such as k-nearest neighbor classifier (k-NN), support vector machine (SVM) and Gaussian mixture classifier (GMC) in terms of 10-fold CV and three-way data splits.

Agent-Based Simulation About Social Value Emergence Based on Perceptual Balance

Hararys structural balance theory based on the idea of Heider explains social processes and is used to account for social actors attitudes toward another. We propose a new social value emergence model in the form of an agent-based simulation model. In this model, the structural balance theory is used to explain the feelings, attitudes and beliefs. Each agent is in effort to reach balanced states and communicates with each other. We analyze a social value emergence by using the proposed model.

State and action space segmentation algorithm in Q-learning

In this paper, we propose a novel Q-learning algorithm that segmentalizes the agent environment and action. This algorithm is learned through interaction with an environment and provides deterministic space segmentation. The purposes of this study can be divided into two main groups: search domain reduction and heuristic space segmentation. In our method, the most activated space segment is divided into new two segments with the learning by a heuristic and recognizable method. Appropriate search domain reduction can minimize the learning time and enables us to recognize the evolutionary process. This segmentation method is also designed for social simulation models. Social space segmentation, such as language systems and culture, is revealed by multi-agent social simulation with our method.

Automatic Adaptive Space Segmentation for Reinforcement Learning

We tested a single pendulum simulation and observed the influence of several situation space segmentation types in reinforcement learning processes in order to propose a new adaptive automation for situation space segmentation. Its segmentation is performed by the Contraction Algorithm and the Cell Division Approach. Also, its automation is performed by “entropy,” which is defined on action values’ distributions. Simulation results were shown to demonstrate the influence and adaptability of the proposed method.

Simple Reinforcement Learning for Small-Memory Agent

In this paper, we propose Simple Reinforcement Learning for a reinforcement learning agent that has small memory. In the real world, learning is difficult because there are an infinite number of states and actions that need a large number of stored memories and learning times. To solve a problem, estimated values are categorized as ``GOOD or ``NO GOOD in the reinforcement learning process. Additionally, the alignment sequence of estimated values is changed because they are regarded as an important sequence themselves. We conducted some simulations and observed the influence of our methods. Several simulation results show no bad influence on learning speed.

Improvement of Particle Filter for Reinforcement Learning

In this paper, we propose a novel framework of learning that uses a particle filter. In a real-world situation, it is difficult to express a continuous state and a continuous action. The problem is solved by using our particle filter, which is one of the methods for dividing a continuous state and a continuous action. Our method needs only a small number of memories and parameters for searching the solution in the space. We conducted pendulum and double-pendulum simulations and observed the difference between the conventional method and the proposed method. Simulation results show there was no bad effect on the received reward.

Proposed particle-filtering method for reinforcement learning

We propose a novel action-search particle-filtering algorithm for reinforcement learning processes. This algorithm is designed to perform search domain reduction and heuristic space segmentation. In this method, each action space is divided into several new segments using particles. Appropriate search domain reduction can minimize learning time and enable the recognition of the evolutionary process of learning. In a numerical experiment, the proposed filtering method is applied to a single-pendulum simulation in order to demonstrate the adaptability of this simulation model.

A perceptual approach to user clustering in collaborative filtering

This paper considers a new approach to user clustering in collaborative filtering. Collaborative filtering is a technique for reducing information overload and is achieved by predicting the applicability of items to user. In neighborhood-based algorithms, the applicability is given by the weighted averages of ratings of neighbors. The new clustering method plays a role for selecting the neighbors based on a perceptual approach, in which users and items are partitioned into two clusters by balancing a general signed graph composed of alternative evaluations on items by users.

Visualization of balancing systems based on naïve psychological approaches

In this paper, we propose a novel medium for interactions based on an interpersonal psychological approach referred to as ‘naïve psychology’. We adopt the visual assessment of clustering tendency (VAT) to naïve psychology for the visual understanding of other people. The VAT algorithm produces a visual display that can be used to assess clustering tendencies in a set of persons (notions) by reconstructing a digital image representation of a square relational dissimilarity matrix for its set. This algorithm clearly represents two types of imbalanced situations in naïve psychology: crisp and fuzzy. The visual image of a balanced or imbalance situation is useful for a deeper human understanding.