Hidekiyo Itakura

Fuzzy network technique for technological forecasting

Abstract The stochastic network technique is known to be a powerful tool carrying out a technological forecast of complex systems. A network dealt with is characterized by a tetrad of essential elements: logical nodes with some inputs and outputs, probabilistics activity branches, feedback loops, and multiple sources and sinks. A set of network parameters is defined for each element and their values are estimated for practical analysis of the network. In the case where the system to be treated is very large and/or complex, it cannot always be represented by a definite network and therefore forecasted values of parameters are inevitably indefinite themselves. A conventional probabilistic approach is sometimes inadequate in such a case. In the light of these facts, the paper proposes a fuzzy network technique, in which among activity branches emanating from a node, a branch to be undertaken once the node is realized belongs to a fuzzy set; and the time required to complete an activity branch belongs to a fuzzy set. Operations of maximum and minimum for sum and product of fuzzy sets take the place of manipulations of addition and multiplication for probabilities, respectively. Although the operations are somewhat formal, the obtained results seem interesting. A numerical example is attached to show a comparison of the proposed technique with the conventional one.

A car detection system using the neocognitron

A car image detection system using the neocognitron is described. The system can recognize car images successfully without regard to influences of the differences of kinds of cars and shifts in position. The number of cell planes can be reduced by actively introducing features of patterns to be recognized by the neocognitron. The neocognitron uses vertical and horizontal lines and combinations as training patterns. The increase of the number of cell planes can thus be held down. Although car images are not directly used in the training process except in the output layer, the system can detect cars skilfully. Thus, using appropriate features of input patterns, the neocognitron obtains sufficient recognition capability.<<ETX>>

An inverse modeling using a five-layer perceptron

This paper shows a learning algorithm for an inverse model of a system using a five-layer perceptron. In the learning algorithm, two performance indexes are used: one is an index for the forward model of the system and the other is for the inverse model. The algorithm reduces these two performance indexes at the same time. As a result, the forward model and the inverse model are formed in the perceptron. The algorithm is applied to the learning of inverse kinematics and dynamics models of manipulators by computer simulations. By the simulation experiments, it is confirmed that the algorithm can learn the inverse models effectively.

A concurrent learning algorithm of forward and inverse models using feedback error learning in the early stage

This paper proposes a concurrent learning algorithm for forward and inverse modeling. The algorithm is consisted of two phases. In the first phase, a feedback controller is used. The forward model is trained using the output values of the controller as the input values to the system and the inverse model is trained by the feedback error learning. In the second phase, the forward model and the inverse model are trained at the same time. By the simulation experiments in a two-link manipulator, it is confirmed that our algorithm can converge faster than the ones already proposed.

A solution to multiple linear regression problems with ordered attributes

Abstract A class of multiple linear regression techniques is discussed, in which the order of magnitude is constrained among regression coefficients. Each predictor variable is a qualitative variate having some categories which are on an ordinal scale. The criterion variable is quantitative. The problem to be solved is reduced to a quadratic programming problem in which the objective function is the residual sum of the squares in regression, and the constraints are linear ones imposed on the regression coefficients. Under some conditions for the observed data, this problem can be solved numerically. The proposed technique works effectively for some types of regression analysis.

An echo canceller with a block-implemented learning identification algorithm

Abstract Characteristics of an echo canceller with use of a block-implemented learning identification algorithm are examined through an approximate analysis and computer simulation experiments. Block digital signal processing, which involves the calculation of a block or finite set of system output from a block of input, is a promising technique to allow efficient use of parallel processors. The echo return loss enhancement, defined as a performance index of the echo canceller, is represented by an approximate equation as a function in the signal-to-noise ratio of a circuit and an acceleration coefficient for updating the echo-path impulse response estimates. This approximate analytical expression is observed to agree well with the results from simulation experiments.