Theses and Dissertations from UMD
Permanent URI for this communityhttp://hdl.handle.net/1903/2
New submissions to the thesis/dissertation collections are added automatically as they are received from the Graduate School. Currently, the Graduate School deposits all theses and dissertations from a given semester after the official graduation date. This means that there may be up to a 4 month delay in the appearance of a give thesis/dissertation in DRUM
More information is available at Theses and Dissertations at University of Maryland Libraries.
Browse
6 results
Search Results
Item Efficient Models and Learning Strategies for Resource-Constrained Systems(2024) Rabbani, Tahseen Wahed Karim; Huang, Furong; Computer Science; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)The last decade has seen sharp improvements in the performance of machine learning (ML) models but at the cost of vastly increased complexity and size of their underlying architectures. Advances in high-performance computing have enabled researchers to train and deploy models composed of hundreds of billions of parameters. However, harvesting the full utility of large models on smaller clients, such as Internet of Things (IoT) devices, without resorting to external hosting will require a significant reduction of parameters and faster, cheaper inference. In addition to augmenting IoT, efficient models and learning paradigms can reduce energy consumption, encourage technological equity, and are well-suited for deployment in real-world applications that require fast response in low-resource settings. To address these challenges, we introduce multiple, novel strategies for (1) reducing the scale of deep neural networks and (2) faster learning. For the size problem (1), we leverage tools such as tensorization, randomized projections, and locality-sensitive hashing to train on reduced representations of large models without sacrificing performance. For learning efficiency (2), we develop algorithms for cheaper forward passes, accelerated PCA, and asynchronous gradient descent. Several of these methods are tailored for federated learning (FL), a private, distributed learning paradigm where data is decentralized among resource-constrained edge clients. We are exclusively concerned with improving efficiency during training -- our techniques do not process pre-trained models or require a device to train over an architecture in its full entirety.Item Characterization and Modeling of Brushless DC Motors and Electronic Speed Controllers with a Dynamometer(2019) Brown, Robert; Chopra, Inderjit; Aerospace Engineering; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)The global drone market is expected to grow from $4.9 billion to $14.3 billion within the next decade, indicating a heavy demand for high performance electric aircraft. Modern drones are propelled with brushless DC (BLDC) motors and electronic speed controllers (ESCs). However, a current lack of information concerning the performance and efficiency of BLDC motors and ESCs prevents their use in rigorous aircraft design. Low cost hobby ESCs and BLDCs are typically used in research aircraft, but few technical details are released by their manufacturers. To better understand these devices, a custom dynamometer was constructed to study the performance of ESCs and BLDC motors. By properly recording the DC, AC, and mechanical power, information on peak efficiency and performance for the ESCs and BLDC motors are determined experimentally. Motors between 920 KV to 2500 KV were tested with 18 A, 30 A, and 40 A ESCs. A combination of these tests were carried out at 7.2 V, 11.1 V, and 14.8 V DC to explore trade offs in the design process. While typically neglected in formal analysis, this work seeks to better understand the power loss mechanisms in ESCs, as it was found that ESCs could have efficiencies as low as 65%, reducing the overall efficiency of the system considerably. This custom dynamometer features a load varying device, power analyzers, and a unique two DAQ setup to properly capture the high frequency electrical signals of BLDC motors. From the sets of experimentally recorded motor and ESC tests, a novel analytical model is developed to predict the performance of ESCs and BLDC motors. At the heart of this modeling effort is describing the 3 phase AC circuit as a single equivalent circuit, which encapsulating the motor’s performance. This work is critical in the design process, as properly sizing ESCs, motors, and rotors for an electric aircraft can improve aircraft endurance and range. Performance metrics are extracted from experimental results and are fit into the analytical model. Predictions for the system’s mechanical power, AC power, and DC power agree well with experimental results, demonstrating applicability of the robust model.Item Essays on Auction Design(2018) Yan, Haomin; Ausubel, Lawrence M; Economics; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)This dissertation studies the design of auction markets where bidders are uncertain of their own values at the time of bidding. A bidder's value may depend on other bidders' private information, on total quantity of items allocated in the auction, or on the auctioneer's private information. Chapter 1 provides a brief introduction to auction theory and summarizes the main contribution of each following chapter. Chapter 2 of this dissertation extends the theoretical study of position auctions to an interdependent values model in which each bidder's value depends on its opponents' information as well as its own information. I characterize the equilibria of three standard position auctions under this information structure, including the Generalized Second Price (GSP) auctions, Vickrey-Clarke-Groves (VCG) auctions, and the Generalized English Auctions (GEA). I first show that both GSP and VCG auctions are neither efficient nor optimal under interdependent values. Then I propose a modification of these two auctions by allowing bidders to condition their bids on positions to implement efficiency. I show that the modified auctions proposed in this chapter are not only efficient, but also maximize the search engine's revenue. While the uncertainty of each bidder about its own value comes from the presence of common component in bidders’ ex-post values in an interdependent values model, bidders can be uncertain about their values when their values depend on the entire allocation of the auction and when their values depend on the auctioneer's private information. Chapter 3 of this dissertation studies the design of efficient auctions and optimal auctions in a license auction market where bidders care about the total quantity of items allocated in the auction. I show that the standard uniform-price auction and the ascending clock auction are inefficient when the total supply needs to be endogenously determined within the auction. Then I construct a multi-dimensional uniform-price auction and a Walrasian clock auction that can implement efficiency in a dominant strategy equilibrium under surplus-maximizing reserve prices and achieve optimal revenue under revenue-maximizing reserve prices. Chapter 4 of this dissertation analyzes an auctioneer's optimal information provision strategy in a procurement auction in which the auctioneer has private preference over bidders' non-price characteristics and bidders invest in cost-reducing investments before entering the auction. I show that providing more information about the auctioneer's valuation over bidders' non-price characteristics encourages those favored bidders to invest more and expand the distribution of values in the auction. Concealment is the optimal information provision policy when there are two suppliers.Item THE INFLUENCE OF CONSCIOUS CONTROL OF MOVEMENT ON BRAIN PROCESSES AND THE QUALITY OF COGNITIVE-MOTOR PERFORMANCE(2015) Lo, Li-Chuan; Hatfield, Bradley D.; Kinesiology; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)The impact of mental stress on fine motor performance is typically maladaptive. The current research was conducted to investigate the manner by which state anxiety affects performance using a cognitive neuroscience perspective. The basic proposition tested, derived from the Reinvestment Theory and the Psychomotor Efficiency Hypothesis, is that stress introduces neuromotor noise to motor planning processes that translate as excess recruitment of motor units and degrade performance. Electroencephalography (EEG) was employed in Study 1 to assess regional cortical activation and cortico-cortical communication between non-motor associative and motor planning regions during the preparatory period of a dart-throwing task. The task was performed during stress (i.e., social evaluation, monetary incentives, and threat of electrical shock) and a relatively relaxed control condition through a within-subjects design. Regional activation was estimated from bilateral EEG recordings in the frontal, central, temporal, parietal, and occipital regions via spectral analysis to assess low-alpha and high-alpha band power to determine generalized arousal and task-relevant attentional focus, respectively. Cortico-cortical communication was estimated between all bilateral regions and the frontal motor planning area with particular emphasis on the left temporal (T3) to midline frontal (Fz) coherence. Elevated state anxiety was induced and associated with heightened T3-Fz EEG connectivity and synchrony of high-alpha band in the right occipital region. Based on these findings, Study 2 was conducted to determine the psychological processes accounting for the observed elevation in T3-Fz EEG coherence and the quality of muscle action during the throwing task. Specifically, participants employed an internal and an external attentional focus to perform the throwing task while their EEG and electromyography (EMG) were monitored. The use of internal focus, which is consistent with explicit monitoring of movement mechanics, was predicted to result in elevated T3-Fz EEG connectivity. This prediction was supported and, furthermore, the magnitude of connectivity was positively associated with motor unit activity assessed via EMG of four major muscle groups (i.e., flexor carpi ulnaris, extensor carpi radialis, biceps brachii, and triceps brachii). The evidence provided supports the theoretical notion that explicit monitoring promotes inefficient muscle activity, which mediates to impact performance negatively.Item Development and analysis of micro polygeneration systems and adsorption chillers(2012) Gluesenkamp, Kyle; Radermacher, Reinhard; Mechanical Engineering; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)About a fifth of all primary energy in the US is consumed by residential buildings, mostly for cooling, heating and to provide electricity. Furthermore, retrofits are essential to reducing this consumption, since the buildings that exist today will comprise over half of those in use in 2050. Residential combined heat and power (or micro CHP, defined by <5 kW electrical generation capacity) has been identified as a retrofit technology which can reduce energy consumption in existing homes during the heating season by 5-30%. This thesis investigates the addition of a thermally-driven chiller/heat pump to a CHP system (to form a trigeneration system) to additionally provide savings during the cooling season, and enhance heating season savings. Scenarios are identified in which adding thermally-driven equipment to a micro CHP system reduces primary energy consumption, through analytical and experimental investigations. The experimental focus is on adsorption heat pump systems, which are capable of being used with the CHP engines (prime movers) that are already widely deployed. The analytical analysis identifies energy saving potential off-grid for today's prime movers, with potential on-grid for various fuel cell technologies. A novel dynamic test facility was developed to measure real-world residential trigeneration system performance using a prototype adsorption chiller. The chiller was designed and constructed for this thesis and was driven by waste heat from a commercially available natural gas-fueled 4 kW (electric) CHP engine. A control strategy for the chiller was developed, enabling a 5-day experiment to be run using a thermal load profile based on moderate Maryland summer air conditioning loads and typical single-family domestic hot water demand, with experimental results in agreement with models. In this summer mode, depending on electrical loads, the trigeneration system used up to 36% less fuel than off-grid separate generation and up to 29% less fuel than off-grid CHP without thermally driven cooling. However, compared to on-grid separate generation, the experimental facility used 16% more primary energy. Despite high chiller performance relative to its thermodynamic limit, this result is primarily due to the electrical efficiency of the prime mover being lower than the grid. A residential trigeneration system utilizing a high temperature fuel cell is predicted to save up to 42% primary energy relative to the grid.Item Is It More Profitable to Post Prices? - Market Structure with Endogeneous Search Costs(2006-08-07) Gong, Binglin; Rust, John; Economics; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)This dissertation contains three chapters. It analyzes a market where firms can choose whether or not to publicly post their prices. Price posting rewards a firm by reducing search costs for customers and thus attracting more demand, at the risk of triggering more direct price competition. In the first two chapters, I use a continuous model and a discrete model to discuss possible market equilibria, respectively. In a non-cooperative and dynamic environment, I find that when the supporting information technology becomes available to all firms, a firm wants to post prices only when it has appropriate cost advantage over its competitors. A lower cost of posting prices encourages firms to post their prices. Price posting improves market efficiency unless one firm has too much cost advantage. When a more efficient entrant replaces the incumbent price-posting firm, the incumbent wants to hide its prices again. These results explain why in some markets firms or individual traders hesitate to publicly post their prices and some even impose search costs on their prices. In the third chapter, I use a laboratory experiment to show how a market evolves when firms or individual traders endogenously determine the search costs on their prices. In the experiment, human subjects play sellers and the computer calculates demands and profits, assuming consumers behave optimally. I assign costs and demand parameters to subjects and let them choose both their prices and whether or not to publicly post them. I alter the production costs, the fixed cost of posting prices, and the possibility of communication among subjects across treatments to show the effect of these factors on market structure. Experimental results show that one is more likely to post prices when he or she has lower unit cost and when the fixed cost of posting prices becomes lower. Price posting lowers effective prices when communication among subjects is not allowed but raises prices when subjects can communicate with each other.