Search

The foreign exchange (FX) market is long established as the largest and most important global financial market. While a large number of research papers focus on forecasting in the FX market, there are still gaps in the literature. First, very few papers focus on improving the parameter estimation process in the forecasting context. Second, artificial neural networks (ANN) with large sizes have not been applied to FX forecasting with the recently fast-developed GPU techniques. Third, forecasting for trading purposes in the FX market has been limited to either building forecasting models or analysing technical indicators. A combination of ANN forecasting models with technical indicators is rare in the existing literature. The use of more accurate parameter estimation algorithms and GPU techniques also makes the thesis unique in the methodological sense. The thesis uses three types of ANN models, namely GARCH-ANN, large Multilayer Perceptron (MLPNN) and Long Short Term Memory (LSTM), to forecast volatility, the direction of price movements and price patterns in the FX market. Research is conducted at three data frequencies, namely monthly, daily and hourly as the analysis goes from the macro-perspective to the micro-perspective. In the first empirical chapter, a Recursive Simulation Algorithm (RSGA) is proposed for estimating the parameters of a volatility forecasting model using GARCH-ANN. The proposed algorithm significantly improves the stability and accuracy of the estimation process by dealing with the local-optimum and convergence problems. The second empirical chapter utilises a large MLPNN model with GPU implementation to forecast the price direction of different FX pairs, with over 40 macro-economic indicators as input variables. Highly profitable out-of-sample results are observed for some of the currency pairs, which challenges the semi-strong form of the Efficient Market Hypothesis (EMH). Significant efficiency improvement is achieved with the GPU implementation. The third empirical chapter proposes the use of the Relative Strength Indicator (RSI) as a measure of the extent of trend-following and mean-reversion patterns of FX rates. A LSTM modelis utilised to forecast price movement patterns (measured by RSI). The trading strategy based on forecasting results of price movement patterns generates more stable profits than the benchmark Moving Average (MA) or RSI implemented on their own. However, the overall low profitability over time for the four currency pairs fails to challenge the weak-form EMH. Overall, with the novel methodologies and technologies implemented within different models, this thesis finds evidence on some extent of inefficiency of the FX market at lower trading frequency (e.g. monthly) and less inefficiency of the FX market at higher trading frequency (e.g. hourly). One possible explanation is that at higher frequencies, the large number of daily (or higher frequency) traders and high-frequency trading algorithms reduce both the number of mis-pricing opportunities and the length of time that any mis-pricing opportunity may last.

Human beings have a touch and force estimation mechanism beneath their feet. They use this feeling of touch and force to maintain balance, walk, run and perform various agile motions. This paper presents a new sensor platform beneath the humanoid feet, enabled by a pragmatic model based compliant foot design and sensor configuration that mimics the human tactile sensory system for contact force measurement in humanoid robots. Unlike previous force sensor based approaches, the system is defined as a total and sufficient method of Ground Reaction Force (GRF) and Zero Moment Point (ZMP) measurement for balancing and walking using contact force feedback in mid to full sized humanoids. The conventional systems for the GRF and ZMP measurement are made of heavy metallic parts that tend to be bulky and vulnerable to inertial noises upon high acceleration. In addition to low cost and reliable operation, the proposed system can withstand shock and enable agile motion much like humans do with their footpad. The proposed foot is manufactured using state-of-the-art technique with elastomer padding which not only protects the sensors but also acts as a compliance beneath the foot giving integrity in structural design. This composite layer provides compliance and traction for foot collision while the contact surfaces are sampled for pressure distribution which can be mapped into three axis force and ZMP. A single step training process is required to relate the sensor readings to force measurement. The system’s capability of contact force measurement, subsequent ZMP estimation is experimentally verified with the application of appropriate software. Moreover, a simulation study has been conducted via Finite Element Analysis (FEA) of the footpad structure to analyze the proposed footpad structure. The experimental results demonstrate why this can be a major step toward a biomimetic, affordable yet robust contact force and ZMP measurement method for humanoid robots. This work was supported by the Office of Naval Research, Grant N00014-15-1-2128 as part of development of Project SAFFiR (Shipboard Autonomous Firefighting Robot).

Artificial neural networks (ANNs) are parallel architectures for processing information even though they are usually realized on general-purpose digital computers. This research has been focused on the design, analysis and real-time realization of artificial neural networks using programmable analog hardware for control and classification. We have investigated field programmable analog arrays (FPAAs) for realizing artificial neural networks (ANN). Our research results and products include a general theoretical limit on the number of neurons required by an ANN to classify a given number of data points, a design methodology for the efficient use of specific FPAA resources in ANN applications, several multi-chip FPAA implementations of ANNs for classification experiments, several single-chip FPAA implementations of analog PID controllers for an unmanned ground vehicle (UGV), experimental evaluation of FPAA PID controllers with a conventional digital PID controller on a UGV, and finally a single-chip FPAA implementation of a (non-linear) ANN controller for comparison with the previous FPAA PID controller on a UGV. 2 These results are collected as four papers formatted for publication and comprising chapters 3, 4, 5, and 6 of this thesis. The first paper develops our general bound for neural network complexity. The second presents a systematic approach based on the upper bound theory for implementing and simplifying neural network structures in FPAA technology. In the third paper, a FPAA based PID controller was designed and characterized in a path-tracking UGV; some of the results from this report are used as a baseline in the fourth paper. In the fourth paper, a FPAA based ANN controller is designed to control a path-tracking UGV and is investigated analytically and with simulation before its performance was experimentally compared to the previously designed FPAA PID controller regarding speed, stability and robustness. In conclusion, this dissertation focuses on the design, analysis and real-time realization of artificial neural networks. The proposed upper bound for neural network complexity provides guidelines for reducing hardware requirements and applies to any layered ANN approach to classification. It is complemented by the neural network structure simplification method which exploits specific features available in the FPAA technology which we used in our experiments and which we believe possess great potential for future real-time control and classification applications.

The dissertation examines advancements in the methods and techniques used in the field of econometrics. These advancements include: (i) a re-examination of the underlying statistical foundations of statistical models with binary dependent variables. (ii) using feed-forward backpropagation artificial neural networks for modeling dichotomous choice processes, and (iii) the estimation of unconditional demand elasticities using the flexible multistage demand system with asymmetric partitions and fixed effects across time. The first paper re-examines the underlying statistical foundations of statistical models with binary dependent variables using the probabilistic reduction approach. This re-examination leads to the development of the Bernoulli Regression Model, a family of statistical models arising from conditional Bernoulli distributions. The paper provides guidelines for specifying and estimating a Bernoulli Regression Model, as well as, methods for generating and simulating conditional binary choice processes. Finally, the Multinomial Regression Model is presented as a direct extension. The second paper empirically compares the out-of-sample predictive capabilities of artificial neural networks to binary logit and probit models. To facilitate this comparison, the statistical foundations of dichotomous choice models and feed-forward backpropagation artificial neural networks (FFBANNs) are re-evaluated. Using contingent valuation survey data, the paper shows that FFBANNs provide an alternative to the binary logit and probit models with linear index functions. Direct comparisons between the models showed that the FFBANNs performed marginally better than the logit and probit models for a number of within-sample and out-of-sample performance measures, but in the majority of cases these differences were not statistically significant. In addition, guidelines for modeling contingent valuation survey data and techniques for estimating median WTP measures using FFBANNs are examined. The third paper estimates a set of unconditional price and expenditure elasticities for 49 different processed food categories using scanner data and the flexible and symmetric translog (FAST) multistage demand system. Due to the use of panel data and the presence of heterogeneity across time, temporal fixed effects were incorporated into the model. Overall, estimated price elasticities are larger, in absolute terms, than previous estimates. The use of disaggregated product groupings, scanner data, and the estimation of unconditional elasticities likely accounts for these differences.

The goal for this paper and my research is to reduce overall cost associated with electricity use at UC Merced. UC Merced presents itself as a unique opportunity for to model integration and optimization of renewable energy sources. It will be discussed exactly what makes UC Merced unique and how UC Merced has set a path towards higher energy efficiency on a community level. Furthermore, I will discuss difficulties involved with integrating renewable resources and then proceed to analyze techniques for further optimization as UC Merced continues its path towards zero net energy. One of these optimization techniques, genetic algorithms; I will discuss in some detail as it was the technique chose to verify the results of the optimization. The main goal of this study is to determine the effect of moving UC Merced‘s Central Plant load closer to or completely during daylight hours when there is inexpensive (solar) energy available or during the night time when energy pricing is minimum. While it seems logical to shift the cooling load, it has yet to be quantitatively shown that such load shifting would be more cost effective. Genetic algorithm (GA)-based Artificial Neural Network (ANN) models are used for demand and energy production forecasting and then GA based cost optimization is performed to find optimum time window for load shifting. We determined that loading shifting can be beneficial and the associated savings are presented for both summer and winter seasons.

Abstract: This research investigates and proposes methods to be used for the automation of the conceptual design phases of variational mechanisms. It employs the concept of feature-based modeling approaches. A method is proposed for integrating the dimensional synthesis, mechanical design and CAD generation phases with minimal designer intervention. Extended feature definitions are used in this research to create a smooth data transfer platform between different engineering tools and applications. This paper also introduces another method by which a set of dimensional data collected from a family of existing products is used to predict possible solutions for a new design. This method, based on artificial neural networks for training and solution generation, is used with optimization algorithms for the dimensional synthesis of mechanisms. An excavator arm mechanism is used as a case study to demonstrate these methods. The design of this mechanism is carried out based on its digging mode configurations.

Spatial or temporal serial remote sensing images are taking more and more important roles in monitoring, utilizing and analyzing resources. However, a large number of remote sensing images are contaminated by clouds, which cause missing information and, moreover, result in the difficulty of extracting complete information. Traditional resolutions to this problem have limits such as low resolution, data lost or large computational load. In this paper, a method that utilizes Artificial Neural Networks (ANN) interpolator is implemented, which may avoid those problems stated above. For the sake of assessing the performance of ANN interpolator, a small area of forest, mountain, valley and road is clipped off from an ETM+ file. Several “cloud” areas will be manually created to test the ANN model. One band of the image is transformed into ASCII files. In the next step, a K Nearest Neighbor (KNN) searching algorithm is applied on these ASCII files, and k neighbors was found for every pixel in this area. Then an ANN model is built. For each pixel that was contaminated by cloud, its neighbors are used as input information, and the output for this pixel is its predicted DN value. Finally, the output will be restored to a raster file. Root-Mean-Square-Error, Quantile-Quantile Plot, and Error Distribution Map are adopted to assess the performance of this ANN interpolator. Finally, the thesis concludes that activate functions and neighborhood search do not cause significant difference in the output of ANN interpolators, and the interpolation results are globally good but largely biased regionally.