Browsing by Author "FU, HC"

Now showing 1 - 10 of 10

A FUZZY NEURAL-NETWORK FOR RULE ACQUIRING ON FUZZY CONTROL-SYSTEMS
(1995-05-12)
SHANN, JJ
;
FU, HC
;
交大名義發表
;
資訊工程學系
;
National Chiao Tung University
This paper presents a layer-structured fuzzy neural network (FNN) for learning rules of fuzzy-logic control systems. Initially, FNN is constructed to contain all the possible fuzzy rules. We propose a two-phase learning procedure for this network. The first phase is a error-backprop (EBP) training, and the second phase is a rule pruning. Since some functions of the nodes in the FNN have the competitive characteristics, the EBP training will converge quickly. After the training, a pruning process is performed to delete redundant rules for obtaining a concise fuzzy rule base. Simulation results show that for the truck backer-upper control problem, the training phase learns the knowledge of fuzzy rules in several dozen epochs with an error rate of less than 1%. Moreover, the fuzzy rule base generated by the pruning process contains only 14% of the initial fuzzy rules and is identical to the target fuzzy rule base.
A HYBRID NEURAL NETWORK MODEL FOR SOLVING OPTIMIZATION PROBLEMS
(1993-02-01)
SUN, KT
;
FU, HC
;
資訊工程學系
;
Department of Computer Science
In this paper, we propose a hybrid neural network model for solving optimization problems. We first derive an energy function, which contains the constraints and cost criteria of an optimization problem, and we then use the proposed neural network to find the global minimum (or maximum) of the energy function, which corresponds to a solution of the optimization problem. The proposed neural network contains two subnets: a Constraint network and a Goal network. The Constraint network models the constraints of an optimization problem and computes the gradient (updating) value of each neuron such that the energy function monotonically converges to satisfy all constraints of the problem. The Goal network points out the direction of convergence for finding an optimal value for the cost criteria. These two subnets ensure that our neural network finds feasible as well as optimal (or near-optimal) solutions. We use two well-known optimization problems-the Traveling Salesman Problem and the Hamiltonian Cycle Problem-to demonstrate our method. Our hybrid neural network successfully finds 100 % of the feasible and near-optimal solutions for the Traveling Salesman Problem and also successfully discovers solutions to the Hamiltonian Cycle Problem with connection rates of 40% and 50%.
A NEURAL NETWORK APPROACH TO THE TRAFFIC CONTROL PROBLEM IN REVERSE BASE-LINE NETWORKS
(1993)
SUN, KT
;
FU, HC
;
交大名義發表
;
資訊工程學系
;
National Chiao Tung University
A neural network for the traffic control problem applied to reverse baseline networks has been proposed in this paper. This problem has been first represented by an energy function. A neural network is applied for maximizing die energy of the function under the constraints of the reverse baseline network. The number of iteration steps in our neural network is limited by a performed upper bound O(n), where n is the size of an n x n network. The throughputs of our neural network have been shown by the empirical results to be better than the conventional algorithm (modified Bipartite Matching Algorithm) when the packet densities rise higher than 50%.
BACKPROPAGATION LEARNING FOR ACQUIRING FINE KNOWLEDGE OF FUZZY NEURAL NETWORKS
(1993)
SHANN, JJ
;
FU, HC
;
資訊工程學系
;
Department of Computer Science
FEATURE-PRESERVING THINNING ALGORITHM FOR OPTICAL CHARACTER-RECOGNITION
(1995-07-01)
FU, HC
;
CHENG, TS
;
CHIANG, CC
;
ROAN, SM
;
交大名義發表
We propose an improved thinning algorithm. Basically, the algorithm uses a 3 x 3 window to accommodate an eight-neighbor skeleton in each thinning iteration. This algorithm overcomes many thinning problems, such as Y-shaped distortions, spiky skeletons and skeleton shortening, and thus preserves precise features of digital character patterns. By using this thinning algorithm, better structural features of a character pattern can be provided to an optical character recognition system, such that accurate recognition results can be achieved.
FISOFM - FIREARMS IDENTIFICATION BASED ON SOFM MODEL OF NEURAL NETWORK
(1994)
KOU, CY
;
TUNG, CT
;
FU, HC
;
資訊工程學系
;
Department of Computer Science
KNOWLEDGE LEARNING ON FUZZY EXPERT NEURAL NETWORKS
(1994)
FU, HC
;
SHANN, JJ
;
PAO, HT
;
資訊工程學系
;
Department of Computer Science
NEURAL NETS FOR RADIO MORSE CODE RECOGNIZING
(1993)
FU, HC
;
LIN, YY
;
PAO, HT
;
資訊工程學系
;
Department of Computer Science
THE CLASSIFICATION CAPABILITY OF A DYNAMIC THRESHOLD NEURAL-NETWORK
(1994-04-01)
CHIANG, CC
;
FU, HC
;
資訊工程學系
;
Department of Computer Science
This paper proposes a new type of neural network called the Dynamic Threshold Neural Network (DTNN). Through theoretical analysis, we prove that the classification capability of a DTNN can be twice as effective as a conventional sigmoidal multilayer neural network in classification capability. In other words, to successfully learn an arbitrarily given training set, a DTNN may need as little as half the number of free parameters required by a sigmoidal multilayer neural network.
USING MULTITHRESHOLD QUADRATIC SIGMOIDAL NEURONS TO IMPROVE CLASSIFICATION CAPABILITY OF MULTILAYER PERCEPTRONS
(1994-05-01)
CHIANG, CC
;
FU, HC
;
資訊工程學系
;
Department of Computer Science
This letter proposes a new type of neurons called multithreshold quadratic sigmoidal neurons to improve the classification capability of muitilayer neural networks. In cooperation with single-threshold quadratic sigmoidal neurons, the multithreshold quadratic sigmoidal neurons can be used to improve the classification capability of multilayer neural networks by a factor of four compared to committee machines and by a factor of two compared to the conventional sigmoidal multilayer perceptrons.