This paper presents a new global optimization method based on ε-approximation approach and its application to lelarning of multilayer neural networks. It is well known that learning of multilayer neural networks can be viewed as minimization of nonlinear objective function with respect to a set of parameters: -weights and thresholds. Since the objective function to be minimized has generally nonconvexity as well as nonlinearity, conventional back propagation methods with local convergence property frequently converge at one of local minima which are false solutions in a sense of learning. The learning method based on the proposed optimization method converges at global minima, which are true solutions in a sense of learning, with less computation time and higher possibility than the learning method based on conventional ε-approximation approach. The numerical results show the superiority of the proposed global optimization method on learning of multilayer neural networks.
This paper describes a computer system supporting negotiations in group decision to seek a solution, where consensus is incomplete, by carring out an encounter-type simulation. The system is based on the utility theory including utility functions which are modified according to the status given by the values and coefficients of utility functions. A consensus is seeked through the modification of the functions. As a concrete example, we demonstrate the use of the system for group car buying decisions. The results can be summarized as follows. (1) A negotiation rule with the concept of utility theory makes it possible to seize negotiators' utility values in quantity which is changeable. And that contributes to clarifying the process of compromise. (2) Those who implement this encounter-type simulation with computer are able to negotiate experimentally and moreover, to give a try to induce a reasonable consensus objectively. (3) Elements and relations in each set or criteria in the system are shown graphically or as relational data in matrix form so that it works as an interface supporting decision making in negotiations by supplying us with figures and tables required instantly in the display.
The efficient use of energy is becoming more important as emerging problems such as global warming and global sustainable development are drawing more attention recently. A way to enlarge technological domain for efficient energy use would be the systematic search of efficient energy conversion technologies by means of synthesizing a variety of knowledge about energy conversion. This paper proposes a synthesis method of energy conversion technologies from elementary energy conversion processes, using qualitative descriptions of elementary processes as well as quantitative indeces to evaluate conversion efficiency. A computer program writte in C language is developed on an engineering workstation to implement the proposed method. Examples of the synthesis of thermodynamic cycles from elementary processes, including evaporation and condensation of working fluid, are also given to demonstrate the effectiveness of the proposed method.
In many applications, Hopfield neural network for optimization problem falls into local minimum and rarely converges to global minimum. To escape from local minimum, the neuron unit was modified to an oscillatory unit by adding a simple self-feedback circuit. This paper proposes a method for direct energy-value extraction from Hopfield neural network to evaluate the output solution. And an oscillatory neural network is constructed by the combination of the oscillatory unit and the energy-value extraction method. The network can extract many solutions sequentially, and can evaluate the solutions simultaneously. The dynamics of the network is examined by computer simulation. Then a small network is realized with electronic circuit after independent implementations of the oscillatory unit and the energy-value extraction circuit. It is confirmed that the oscillatory neural network is suitable for hardware implementation.
A problem of determination of feedback gains in order to achieve required performances is considered in this paper. We suggest a method that assigns any poles in a sector region which specifies the stability and damping factor by making use of coordinate transformations that is; the imaginary axis's parallel movement and the complex plane's rotation. At first we transform the system to the augmented system by using the above transformation. Next we can show any poles can be located in the specified region very easily by designing the feedback gains which minimize the cost function in the augmented system. Finally it is shown in two numerical examples that our proposed method is effective.
Crossover is one of the most important operators in genetic algorithms, on which the overall performance of the algorithms critically depends. In this paper, we review a variety of crossover operators proposed for sequencing problems, and analyze the relationship between characteristics of the operator and performance of the algorithm. From this analysis, we propose simple criteria for measuring the quality of crossover operators. Some computational analysis on single machine scheduling problem is then added to validate the effectiveness of the proposed criteria.
In this paper we theoretically analyze local minima of a neural network for solving the Travelling Salesman Problem (TSP) and find the neuronal matrix has at most one fired neuron on each row and column. We also find theoretical values of the penalty factor that can make local minima always feasible or always infeasible. Further, we propose a method to improve feasibility of the local minimum: the penalty factor is initially set to a very small value to obtain an infeasible tour with short subtours and, afterwards, is gradually increased finally to obtain a feasible trip. Computational results for the TSPs with random 5_??_30 cities indicate that theoretical results are verified arid the proposed method improve the quality of the solution by 48% on the average compared with the conventional method.