IDS法を用いたモデリングシステムの実用性:性能調査とハードウェア実装

摘要

The performance characteristics robustness, speed, and tractability are important for the realization of practicalcomputing systems. The concept of soft computing enables the achievement of such practical characteristicsby tolerating imprecision and uncertainty instead of depending on exact mathematical computations. Theink drop spread (IDS) method is a modeling technique that has been proposed as a new approach to softcomputing. This method is characterized by a modeling process that uses image information without includingcomplex formulas. The model structure is a parallel distributed system comprising multiple SISO units, eachof which independently functions as a modeling engine; each SISO unit is termed an “IDS unit.” Each IDSunit extracts a feature of the modeling target in the form of an image. The upper processing unit performsinference by aggregating the image information from all the IDS units.The objective of this dissertation is to investigate the performance of the IDS method as a soft computingtool and to reveal the practicality of this method. First, the modeling ability of the IDS method is presentedby using three benchmarks concerning logic operation, function approximation, and classification. Next,the performance of the IDS method is investigated in terms of robustness, speed, and tractability, which aretypical criteria that determine the importance of soft computing tools. The robustness is evaluated on the basisof noise tolerance and fault tolerance. Based on these criteria, we compare the IDS models with artificialneural networks (ANNs) and fuzzy inference systems (FISs).In several studies on ANNs, Hwang’s five-function set is used as the regression benchmark, and the generalizationperformance is evaluated for each set of noiseless and noisy training data. By using this benchmark,we present the noise tolerances of the IDS models, multilayer perceptrons (MLPs), radial basis function networks(RBFNs), and adaptive neuro-fuzzy inference systems (ANFISs). The MLP is the most popular modelof neural networks. The RBFN is a variant of the ANN and is known to have superior fault tolerance andfast convergence. The ANFIS is characterized by its hybrid learning; the parameters of the premise part andconsequent part in its fuzzy inference are adjusted by the gradient descent method and least squares method,respectively.The structure of IDS models is similar to that of ANNs: they comprise distributed processing units. Theimplementation of such parallel processing networks with dedicated hardware provides fault tolerance. Weevaluate the fault tolerances of IDS models and ANNs for single unit failure using the stuck-at fault model.Most studies on the fault tolerance of ANNs investigate fault tolerance using classification tasks. This isbecause in the approximation of continuous functions, the failure of a single hidden unit causes considerabledegradation in computation. Our experiments use function approximations that are more difficult thanclassification for the evaluation of fault tolerance.The speed of learning is important in real-time learning applications that require quick adaptation to nonstationaryconditions. In ANNs, online learning is often used for real-time learning applications; it updatesthe network parameters for every training example and each example is discarded after the update. Onlinelearning in the IDS method also follows this procedure. The real-time learning capabilities of each system areevaluated by using online learning.Since tractability cannot be quantitatively measured and its definition is ambiguous, in this dissertation, itis discussed from the viewpoints of the interpretability and transparency of the model. In the former viewpoint,we follow the definition of interpretability that is generally used. The interpretability is based on a comprehensiblesystem structure and description for humans. On the other hand, the latter viewpoint does not dealwith the transparency for interpretation and analysis, which is the definition of transparency for fuzzy systems;instead, it deals with the transparency with regard to performance. When optimal performance is realized withease for various requirements, we consider such a system to be transparent with regard to performance.Since IDS models have a parallel processing structure, the development of their hardware is beneficialto the improvement of the processing speed. This is evident from the fact that numerous VLSI devices anddedicated processors comprising ANNs and FISs with parallel processing structures have been developed.Moreover, the IDS method is potentially used for fault-tolerant computing. If dedicated IDS hardware units areprovided, the fault tolerance characteristic is realized. This dissertation describes the hardware implementationof the IDS unit. By comparing the processing speed of an IDS modeling system equipped with IDS hardwareunits with that of a software-based IDS modeling system, the effectiveness of the hardware implementation isverified.On the basis of the results of the above evaluations, this dissertation demonstrates that the IDS method isa practical method that has superior capability to function as a soft computing tool.