武藤 佳恭

人工ニューラルネットワークを用いてパーセプトロンや連想メモリをはじめとするさまざまなパターン認識のための技法が提案されてきた。入力パターンをある一つのカテゴリーに分類することはこれらの技法で可能である。しかし複数のパターンが重なっている場合に、それらを的確に分類し認識するためには複雑な構造を必要としてきた。提案手法はカオスニューロンとパターン認識のための古典的な学習法であるバックプロパゲーション技法を組み合わせることにより、これを実現した。非常に扱いやすい単純なモデルにもかかわらず、カオスのダイナミクスを用いることにより、効果的に重なったパターンを分離できる。

In this paper, we propose an algorithm for tracking a rigid object using optical flow. Our algorithm processes optical flow and detects the object's motion in the direction of depth. As a result, three conditions of the moving object; approaching, receding, and static in depth, are determined.A concept of Feature extraction is commonly applied to Object tracking in order to establish a robust detection system against noise. The system based in the proposed algorithm utilizes a perception of a motion in the direction of depth to detect a significant feature for providing the precise detection of a moving object. In the simulation, optical flows are captured by Artificial Retina Camera of MITSUBISHI with 32 * 32 pixel resolution, and the proposed system pursue a moving hand in real time. The simulation result shows that perceiving a motion in the direction of depth improves the performance of conventional object tracking systems.

In this paper, the concept of Boxed Economy model is proposed. Boxed Economy model is an agent-based microeconomic simulation which is intended to make predictions and explanations of the real economic movements in the general economy. We suggest the importance of reusable design with object-oriented framework in the field of agent-based economic simulation. Two frameworks at the different level are described in this paper: "Boxed Simulator Framework" and "Boxed Economy Framework". The source code and model is going to be open to public for sharing and co-improving.

本論文では、自己組織化手法として知られている弾性ネット手法を拡張した弾性テンプレートモデルを用いた手の形状認識手法を提案する。本手法はトップダウンな制約であるテンプレートからの入力とボトムアップな制約である入力画像からの入力を相互作用させることにより画像中から目的のパターンを抽出する。この相互作用の効果により、対象物の位置ずれや歪みを吸収した認識が可能になった。本論文では、数種類の手の形状パターンを用いて、弾性テンプレートモデル、ボトムアップ手法、トップダウン手法の3種類の実験結果を比較し、リアルタイム性と認識精度における本手法の有効性を示した。

本論文で提案するのはニューラルネットワークを用いた対戦型ボードゲームの最適な戦略アルゴリズムである。提案手法の目的は与えられた局面から次の最善の一手を決定することだが、そのために先読みをして未来の両対局者の形勢が互角となる局面を探し出す。この提案手法を用いて五目並べゲームの画期的な戦略を紹介する。本提案手法は、枝分かれにより膨大に派生していく未来の想定局面の中から、目標となる両対局者互角の局面までの道のりを、悪い手を修正しながら探していく。このフィードバック型の戦略は様々なボードゲームのプロのプレイヤーが用いる発見的思考方法をモデル化したものである。プロのプレイヤーの着手と比較したシミュレーションにより、本提案手法が従来手法よりもはるかに少ない計算コストでよい答えを得られることが証明された。

既知不正アクセス手法に関するシグネチャを用いてネットワークトラフィックから不正なアクセスの検出を行う侵入検出システムの開発についてオープンソース開発モデルの適用例を示す.ソースコードとシグネチャの共有によって, 従来不正検出手法を用いた侵入検出システムの問題点とされていたシグネチャ管理やローカライズの問題を解決することができ不正アクセスの動向に応じた効率的な検出が可能となる.

本論文では, ニューラルコンピューティングの組合せ最適化手法の応用として, 現代語を組み合わせた現代風「いろは歌」の作成方法を紹介する.ここでは, 現代風「いろは歌」作成問題を二種類の組み合わせ最適化問題として捉えることによりニューラルコンピューティング手法により解くことを可能にした.一つ目の問題はすべての仮名を重複なく用いるような文節の組を選ぶ問題である。二つ目の問題は日本語の係り受け制約に基づいて語順を決める問題である.シミュレーションでは, あらかじめ語群として用意した964語の文節の中からいろは歌作成を行なった。その結果, 自然な日本語に近い現代風「いろは歌」を生成できることが示された.

A near-optimum parallel algorithm for solving the one-dimensional gate assignment problem is presented in this paper, where the problem is NP-hard and one of the most fundamental layout problems in VLSI design. The proposed system is composed of n×n processing elements based on the artificial two-dimensional maximum neural network for (n+2)-gate assignment problems. Our algorithm has discovered improved solutions in the benchmark problems compared with the best existing algorithms. The proposed approach is applicable to other VLSI layout problems such as the PLA (Programmable Logic Array) folding problem.

This paper proposes a neural network system to detect human faces. Our scheme is composed of a preprocess and KenzanNET. Preprocessing analyzes hyperspectral images by using a hybrid self-organizing classification model to extract skin area and making a facial candidate pattern based on the extracted skin area. KenzanNET discriminate a face from other body parts. KenzanNET is a kind of feed forward neural network and is made from CombNET [ I[ improved by an additional learning function. Under the various conditions in terms of background and brightness in a room and the distance between people and camera, our system can detect human face with 76.9% accuracy.

In this paper, a new neural computing method to extract logical rules from the training data sets is proposed. Maximum neural networks are used to train the weight and the threshold of the multi-layered (feedforward) neural network (MLNN). The threshold and the weights of the MLNN are trained to be a logical function (AND/OR) with the multiple input. The maximum neural network constructs the logical function on the MLNN so that it is not necessary to extract rules from the trained MLNN. The proposed method was experimented for the classification problem, Monk's problem 1. Experimental results showed that the proposed method learned the correct rule in more than 40% success rate.

The vehicle routing problem (VRP) is one of the well known optimization problems. It is used to minimize the total length of all routes of vehicles where each of the vehicle has a capacity constraint respectively. This paper proposes a self-organization neural network model for obtaining the best solution for VRP. Our method consists of two phases. In the first phase, the customers are grouped to several delivery areas for vehicles assignment by Maximum Neuron model. In the second phase, the TSP in each area is solved by Elastic net model proposed by Andrew et. al. The clustering algorithm used in the first phase is a Maximum Neuron model. Maximum Neuron model is one of the neural networks proposed by Hopfield that can minimize a cost function considering various constraints. In the second phase, Elastic net model is used to solve the problem and it can obtain good solutions of TSP. Our method improves the precision of solution, and can be extended for big size problem. Our simulation result shows that Maximum Neuron model can achieve better solutions than other methods in certain conditions.

In this paper, we propose a multi-module chaotic associative memory (MCAM) that uses chaotic neural networks. In this method, the chaotic associative memories are connected to each other. If MCAM can not obtain enough information of a target, MCAM shows a behavior that looks like human "perplexity" where MCAM succeeds in one-to-many associations. And when MCAM obtains enough information to recognize a target, MCAM converges to a stable state. Although the structure of MCAM is simple, MCAM realizes one-to-many association by using chaotic dynamics.

In this paper, we propose a neural network system for verifying whether a mouth or eyes can be extracted from an image area by Back Propagation (BP). It is necessary to test the proposed system in a noisy environment. In this paper, the model of neural network system for recognizing a mouth is based on the function of peripheral vision, In our research, a mouth has distinct properties of brightness in the right corner of the mouth, the left corner of the mouth, the tip of nose, and the nostril. Furthermore we discovered that humans commonly observe these properties of the mouth regardless of the brightness of lighting, different colors of the mouth, or different form of the mouth. By using these features, we designed an associative memory neural network for the verification.

A novel approach to the one-dimensional gate assignment problem is presented in this paper where the problem is NP-hard and one of the most fundamental layout problems in VLSI design. The proposed system is composed of n×n processing elements called the artificial two-dimensional maximum neurons for (n+2)-gate assignment problems. We have discovered the improved solutions in the benchmark problems over the best existing algorithms. The proposed parallel algorithm is also applicable to other VLSI layout problems.

Microcode optimization is an NP-complete combinatorial optimization problem. This paper proposes a new method based on the Hopfield neural network for optimizing the wordwidth in the control memory of a microprogrammed digital computer. We present two methodologies, viz., the maximum clique approach, and a cost function based method to minimize an objective function. The maximum clique approach albeit being near O(1) in complexity, is limited in its use for small problem sizes, since it only partitions the data based on the compatibility between the microoperations, and does not minimize the cost function. We thereby use this approach to condition the data initially (to form compatibility classes), and then use the proposed second method to optimize on the cost function. The latter method is then able to discover better solutions than other schemes for the benchmark data set.

In this article we examine our remote lecture experiments using ATM and UNII wireless devices. We discussed which functions should be improved in the current distance-learning system. UNII/SUPERNet wireless devices are introduced for achieving economical distance-learning systems. Two experiments of remote lectures are detailed: a wireless experiment between two buildings in the Philippines and an ATM experiment between Japan and the United States.

In this paper, the agent-based artificial market model is proposed as a simple model to describe the society as a multi-agent system. We examine the case study of competition on the standardization of video cassette recorders (VCR). In this case the compatibility of the video software is definitely important to choose the VCR. The simulation result shows that the communications among agents are breaking the symmetry of the market share, then the asymmetry is enlarged further with accelerated speed. The model is useful to understand what has happened in the technological lock-in phenomena.

The purpose of this study is to propose the adaptive agent as hybrid of Genetic Algorithm and Neural Network, and to clarify the effectiveness of the combination of two mechanisms in the dynamic environment. Evolution and learning can be explained as the mechanism of searching a solution in the enormous possibilities at the population level and individual level respectively. Genetic algorithm and neural network are computational models. Genetic algorithm is suitable for global search, and neural network at the local search. Combination of genetic algorithm and neural network seems natural from the biological viewpoint. There are two ways of combination of genetic algorithm and neural network, that is Darwinian and Lamarckian framework. In Lamarckian framework the acquired traits during the lifetime can be passed on to the offspring directly, and in Darwinian framework, these cannot be passed on. We propose `Neural Agent' whose initial weights of their neural networks are determined by their genome data, as a simple model of hybrid system of genetic algorithm and neural network. We examine which framework is better in the dynamic system. The result of our simulation shows Darwinian framework is better than Lamarckian.

This paper presents a new self-organization classification algorithm for remote-sensing images. Kohonen and other scholars have proposed self-organization algorithms. Kohonen's model easily converges to the local minimum by tuning the elaborate parameters. In addition to others, S.C. Amatur and Y. Takefuji have also proposed self-organization algorithm model. In their algorithm, the maximum neuron model (winner-take-all neuron model) is used where the parameter-tuning is not needed. The algorithm is able to shorten the computation time without a burden on the parameter-tuning. However, their model has a tendency to converge to the local minimum easily. To remove these obstacles produced by the two algorithms, we have proposed a new self-organization algorithm where these two algorithms are fused such that the advantages of the two algorithms are combined. The number of required neurons is the number of pixels multiplied by the number of clusters. The algorithm is composed of two stages: in the first stage we use the maximum self-organization algorithm until the state of the system converges to the local-minimum, then, the Kohonen self-organization algorithm is used in the last stage in order to improve the solution quality by escaping from the local minimum of the first stage. We have simulated a LANDSAT-TM image data with 500 pixel × 100 pixel image and 8-bit gray scaled. The results justifies all our claims to the proposed algorithm.

The graph partitioning problem is a famous combinatorial problem and has many applications including VLSI circuit design, task allocation in distributed computer systems and so on. In this paper, a novel neural network for the m-way graph partitioning problem is proposed where the maximum neuron model is used. The unidirected graph with weighted nodes and weighted edges is partitioned into several subsets. The objective of partitioning is to minimize the sum of weights on cut edges with keeping the size of each subset balanced. The proposed algorithm was compared with the genetic algorithm. The experimental result shows that the proposed neural network is better or comparable with the other existing methods for solving the m-way graph partitioning problem in terms of the computation time and the solution quality.