Norio Baba

A new approach for finding the global minimum of error function * of neural networks

Abstract Back-propagation may be the most widely-used method to adapt artificial neural networks for pattern classification. However, an important limitation of this method is that it sometimes falls into a local minimum of the error function. In this paper, the random optimization method of Matyas and its modified algorithm are used to learn the weights and parameters in a neural network. Research indicates that these algorithms can be successfully utilized in order to find the global minimum of error function of neural networks.

A modified convergence theorem for a random optimization method

Abstract The purpose of this paper is to investigate the random optimization method which has been contrived for the minimization method. Concerning with the convergence of this method, Matyas gave interesting theorems. However, there is a questionable point in his proofs, and the assumptions used in them are too severe. In this paper, a modified theorem concerned with the convergence of the random optimization method is given.

On the Learning Behavior of Stochastic Automata Under a Nonstationary Random Environment

We propose a new nonstationary random environment R(C1(t,¿),...,Cr(t,¿)), where t represents time and ¿ ¿ ¿, ¿ being the supporting set of a probability measure space (¿,B,¿). Moreover, the learning performance of the Lr-1 scheme under R(C1(t,¿),..., Cr(t,¿)) is discussed.

An application of the stochastic automaton to the investment game

Abstract An application of the stochastic automaton to the investment game is considered. It is shown that the use of the stochastic automaton with learning properties is an efficient method for the investment game.

Learning behaviors of hierarchical structure stochastic automata operating in a general multiteacher environment

The learning behaviours of hierarchical structure automata operating in general multiteacher environments are considered. It is shown that the generalized version of the reinforcement algorithm is absolutely expedient.

Interactive multi-objective programming technique using random optimization method

An interactive optimization algorithm for multi-objective programming that utilizes the random optimization method is proposed. It is shown that a personal computer can be used successfully in order to find an appropriate solution in a dialogue mode.

On the learning behaviors of variable-structure stochastic automaton in the general n-teacher environment

Learning behaviors of a stochastic automaton operating in a multiteacher environment are considered. As a generalized form of the <i>L</i>_R-I scheme, the <i>GL</i>_R-I scheme is proposed as a reinforcement scheme in a multiteacher environment. It is shown that the <i>GL</i>_R-I scheme is absolutely expedient and ϵ-optimal in the general <i>n</i>-teacher environment. Learning behaviors of the <i>GL</i>_R-I scheme are simulated by computer and the results indicate the effectiveness of the <i>GL</i>_R-I scheme.

Learning behaviors of the hierarchical structure stochastic automata under the nonstationary multiteacher environments and their applications to intelligent robot manipulators

Learning behaviors of hierarchical structure stochastic automata are considered in a nonstationary multiteacher environment. It is shown that an extended form of an algorithm proposed by M.A.L. Thathachar and K.R. Ramakrishnan (1981) ensures absolute expediency under some conditions. As a practical application of hierarchical structure stochastic automata, intelligent behavior is considered of robot manipulators going through a maze having a large number of gates that close with unknown rejecting probabilities. It is shown that hierarchical structure stochastic automata can be successfully utilized to let robot manipulators find the best way through the maze.

The absolutely expedient nonlinear reinforcement schemes under the unknown multiteacher environment

Learning behaviours of variable-structure stochastic automata under a multiteacher environment are considered. The concepts of absolute expediency and ε-optimality in a single-teacher environment are extended by the introduction of an average weighted reward and are redefined for a multiteacher environment. As an extended form of the absolutely expedient learning algorithm, a general class of nonlinear learning algorithm, called the GAE scheme, is proposed as a reinforcement scheme in a multiteacher environment. It is shown that the GAE scheme is absolutely expedient and ε-optimal in the general n-teacher environment. Learning behaviours of the GAE scheme in various multiteacher environments are simulated by computer and the results indicate the effectiveness of the GAE scheme.

Intelligent collision avoidance planning of a robot manipulator whose workspace includes moving objects

The collision avoidance problem of a robot manipulator whose workspace includes moving objects is considered in this paper. It is shown that the proposed computer simulation system can be used in a dialogue mode with a designer to check whether or not collision with obstacles is avoided and to determine the appropriate movement.