| 摘要 |
PROBLEM TO BE SOLVED: To stabilize the operation as a finite state machine by providing a term for the scale of pertinence degree of internal state representation in an evaluation function to be optimized and performing maximum likelihood estimating for the coefficients of respective terms. SOLUTION: An initialization part 1 of a neural network initializes a combined matrix with a random number and sets a hyperparameter to an initial value, and a minimization part 2 for T(W) minimizes the evaluation function T(W). An update part 3 for the hyperparameter updates the hyperparameter and an output value calculation part 4 for a neuron finds the output value of the neuron. A square sum calculation part 5 for the combined matrix calculates the sum of squares of the combined matrix and a square error calculation part 6 finds the sum of squares of a tutor signal and an error of the neural network. A calculation part 7 for internal state representation calculates an evaluation term of the internal state representation, a function minimization part 8 finds the minimum value of the function, and an update part 9 for the combined matrix varies values of the combined matrix. In this case, a term of a proper scale of internal state representation is provided for the evaluation function to be evaluated in addition to terms including an error term and a penalty term and the maximum likelihood estimation is performed for the coefficients of the respective terms.
|
