An ad-hoc network operates without a central entity or infrastructure, and is composed of highly mobile hosts. In ad-hoc network, routing protocols are with host mobility and bandwidth constraints. There have been many recent proposals of routing protocols for ad-hoc networks. A recent trend in ad hoc network routing is the reactive on-demand philosophy where routes are established only when required. AODV(Ad-hoc On-demand Distance Vector routing) evaluates routes only on an as-needed basis and routes are maintained only as long as they are necessary. Because the network topology changes frequently in ad-hoc networks, some on-demand protocols with multi-paths or backup routes have been proposed. Sung-Ju Lee and Mario Gerla proposed an AODV-BR scheme to improve existing on-demand routing protocols by creating a mesh and providing multiple alternate routes . The algorithm establishes the mesh and multi-path using the RREP (Route Reply) of AODV, which does not transmit any extra control message. In this paper, we propose two schemes : AODV-ABR(Adaptive Backup Route) and AODV-ABL (Adaptive Backup Route and Local repair) to increase the adaptation of routing protocol to topology changes by modifying AODV-BR. In AODV-ABR, the alternative route can be creating by overhearing not only RREP packets but also data packets. AODV-ABL combines the benefits of AODV-ABR and Local Repair. Finally, we evaluate the performance improvement by simulation.

Mass Rapid Transit(MRT) plays a very important role for the city development,the investment cost is very expensive. It is necessary to derive the MRT system planning by considering the service reliability and performance index according to the forecast of annual ridership. With the less ridership as compared to Taipei MRT network, Kaohsiung MRT has to be developed to achieve the most cost effective investment of power supply and rolling stock planning. This thesis is to investigate the proper expansion planning of traction substations (TSS) for an electrified mass rapid transit system. The motion equation of train sets is used to solve the mechanical power consumption at each time snapshot according to the operation timetable, the passenger ridership and various types of operation resistance. The mathematical models of power converters in traction substations for different operation modes have been derived. With all train sets operated along the main line, the AC/DC load flow analysis is performed to find power demand of all traction substations for annual system peak operation over the study period. The objective function is formulated by considering both the voltage drop of train sets and investment cost of traction substations as the equivalent cost of all feasible states of each year. By performing the dynamic programming (DP) and immune algorithm (IA), the expansion planning of traction substations to achieve the minimum overall cost has been solved by identifying the optimal capacity and locations of new traction substations to be committed at each year.

In this thesis, we introduce a signal-to-noise ratio (SNR) estimator and analyze the performance degradation of the SNR estimator due to the synchronization error in the orthogonal frequency division multiplexing (OFDM) systems. This SNR estimator through the expectation-maximization (EM) algorithm is used in the adaptive modulation. When the synchronization is imperfectly done, the synchronization error reduces the performance of the OFDM systems and the accuracy of the SNR estimator. We investigate the estimation offset of the SNR estimator with the synchronization error. Simulation results demonstrate that the theoretical analyses are correct. In addition, the simulation results show that the more synchronization errors cause the more estimation errors of the SNR estimator. And the estimation errors are not decreased by the iteration of the EM algorithm.

This thesis is based on the DSP to develop a facial expression recognition system. Most facial expression recognition systems suppose that human faces have been found, or the background colors are simple, or the facial feature points are extracted manually. Only few recognition systems are automatic and complete. This thesis is a complete facial expression system. Images are captured by CCD camera. DSP locates the human face, extracts the facial feature points and recognizes the facial expression automatically. The recognition system is divided into four sub-system: Image capture system, Genetic Algorithm human face location system, Facial feature points extraction system, Fuzzy logic facial expression recognition system. Image capture system is using CCD camera to capture the facial expression image which will be recognized in any background, and transmitting the image data to SRAM on DSP through the PPI interface on DSP. Human face location system is using genetic algorithm to find the human face¡¦s position in image by facial skin color and ellipse information, no matter what the size of the human face or the background is simple. Feature points extraction system is finding 16 facial feature points in located human face by many image process skills. Facial expression recognition system is analyzing facial action units by 16 feature points and making them fuzzily. Judging the four facial expression: happiness, anger, surprise and neutral, by fuzzy rule bases.. According to the results of the experiment. The facial expression system has nice performance on recognition rate and recognition speed.

A nonparametric test for the interval-censored failure time data is proposed in determining whether p lifetime populations come from the same distribution. For the comparison problem based on interval-censored failure time data, Sun proposed some nonparametric test procedures in recent year. In this paper, we present simulation procedures to verify the test proposed by Sun. The simulation results indicate that the proposed test is not approximately Chisquare distribution with degrees of freedom p-1 but Chisquare distribution with degrees of freedom p-1 times a constant.

The thesis illustrates the development of DSP-based¡§Real-Time Human Face Recognition System¡¨.The principal system consists of three major subsystems.There are Image Acquisition System¡AImage Preprocessing System and human face characteristic extraction . For experiment,we adopted colored face image with complex background and simulate on PC.We found the characteristic points and characteristic vectors from the face image which is searched from Gene algorithm.Then,we use the recognition system to recognize the face image.Finally we implant it to DSP. Shown by the experimental result,this system has good recognition and efficiency.

The bid-ask spread costs consist of three components, which include order processing costs, inventory-holding costs, and adverse selection costs. In this paper, we model the inventory-holding costs of the bid-ask spread by option hedging. Theinventory-holding costs are hedged by call or put option positions. Since trades deal with the adverse selection traders are unobservable. We treat it as a latent variable, and Expected-Maximization (EM) algorithm are applied to estimate the related parameters of the model. Simulation studies are performed for several different models. Empirical results of NYSE high frequency data show that the proposed model are obtain appropriate parameter estimation when the returns satisfied normality assumption.

DNA microarrays are widely used techniques in molecular biology and DNA computing area. Before performing the microarray experiment, a set of subsequences of DNA called probes which are complementary to the target genes of interest must be found. And its reliability seriously depends on the quality of the probe sequences. Therefore, one must carefully choose the probe set in target sequences. A new method for probe design strategy using multi-objective genetic algorithm is proposed. The proposed algorithm is able to find a set of suitable probes more efficient and uses a model based on suffix tree to speed up the specificity constraint checking. The dry dock experimental results show that the proposed algorithm finds several probes for DNA microarray that not only obey the design properties, but also have specificity.

In this dissertation, we focus on the study of the partial sorting (generalized sorting) problem and the initialization problem. The partial sorting problem is to find the first k smallest (or largest) elements among n input elements and to report them in nondecreasing (or nonincreasing). The initialization problem on a multiprocessor system is to assign each of n input elements a unique identification number, from 1 to n. This problem can be regarded as a special case of the sorting problem in which all input elements have the same value. We propose some algorithms for solving these problems. The main result is to give precise analysis for these algorithms. On the traditional model, we modify two algorithms, based on insertion sort and quicksort, to solve the partial sorting problem. Our analysis figures out the whole race between the two partial sorting algorithms and shows that the partial insertion sort algorithm obtains the leading position from k = 1 (the beginning) until k 3 5pn. After that, the partial quicksort algorithm will take the leading position on the way to the end. We also extend the partial sorting problem on the Single-Hop wireless network with collision detection (WNCD) model. The extension fits in with the wireless trend and may be a foundation for studying divide-and-conquer. With the repeat maximum finding scheme, we propose a partial sorting algorithm and prove that its average time complexity is (k + log (n − k)). For the initialization problem on the WNCD model, we can invoke the sorting algorithms directly for solving it. However, those sorting algorithms would not be better than the method of building a partition tree. We show that the partition tree method requires 2.88n time slots in average. After reconstructing and analyzing the method, we improve the result from 2.88n to 2.46n.

The implicit function model is nowadays generally applied to a lot of fields that need 3D, such as computer game, cartoon or for specially effect film. So far, most hardware are still to support the polygon-mesh model but not implicit function model, so polygon-mesh model is still the mainstream of computer graphics. However, translation between the two representation models becomes a new research topic. This paper presents a new method to translate the triangles mesh model into the implicit functions model. The main concept is to use the binary space-partitioning tree to divide the points and patches in the triangle mesh model to create a hierarchical structure. For each leaf node in this hierarchical structure, we would generate a corresponding implicit function. These implicit functions are generated by the genetic algorithm. And the internal nodes in this hierarchical structure are blended by the blending operators. The blending operators make the surface become smooth and continual. The method we proposed reduces the data in a large amount because we only save the coefficients of the implicit surface. And the genetic algorithm can avoid the high computing complexity.