Electrical Engineering and Computer Sciences - Ph.D. / Sc.D.
http://hdl.handle.net/1721.1/7660
2016-12-06T16:06:28ZOptimization problems with incomplete information
http://hdl.handle.net/1721.1/105676
Optimization problems with incomplete information
Hwang, Daw-sen
This thesis is concerned with the design and analysis of algorithms for new variants of optimization problems where the problem instance is not completely known. Specifically, we consider two online problems where the problem instance is revealed over time, and one distributed problem involving many computational units, each of which can access only local information. We measure the performance of algorithms by the worst-case ratio between their objective values and the optimal objective value obtained by algorithms knowing the entire problem instance. Better algorithms have ratios closer to one. For online problems, this ratio is known as the competitive ratio. First, we study a class of generalized online scheduling problems, where the online Weighted Traveling Repairman Problem (WTRP) is a special case. For the online WTRP, we propose a family of parameterized deterministic and randomized online algorithms. For both deterministic and randomized cases, our competitive ratios are the smallest (best) in the literature. For the general setting, our online algorithms achieve similar competitive ratios. Second, we study a distributed version of the Multi-Depot Vehicle Routing Problem. In particular, we divide the space into smaller regions based on the depot configurations through a partition scheme, and assign each region to a vehicle. For partition schemes, we call the aforementioned worst-case ratio their Price of No-Communication (PoNC). We show that the Voronoi partition achieves a PoNC linear in the number of depots. In addition, for two special classes of depot configurations, we design partition schemes with PoNCs sub-linear in the number of depots. Third, we study quantity-based single-resource revenue management problems in a new parameterized online model, which is a combination of the worst-case and the random-order models. When there are only two classes of customers, we develop two online algorithms and show that they achieve the best-possible competitive ratio for a wide range of problem parameters. We also study two problem extensions. In the first extension, online algorithms can observe whether an arriving customer follows the adversarial arrival order. In the second extension, similar to the classical secretary problem, the goal of the online algorithms is to maximize the probability of selecting the highest-valued customer.
Thesis: Ph. D., Massachusetts Institute of Technology, Department of Electrical Engineering and Computer Science, 2016.; Cataloged from PDF version of thesis.; Includes bibliographical references (pages 277-287).
2016-01-01T00:00:00ZContinuous representations and models from random walk diffusion limits
http://hdl.handle.net/1721.1/105670
Continuous representations and models from random walk diffusion limits
Hashimoto, Tatsunori B. (Tatsunori Benjamin)
Structured data such as sequences and networks pose substantial difficulty for traditional statistical theory which has focused on data drawn independently from a vector space. A popular and empirically effective technique for dealing with such data is to map elements of the data to a vector space and to operate over the embedding as a summary statistic. Such a vector representation of discrete objects is known as a 'continuous representation'. Continuous space models of words, objects, and signals have become ubiquitous tools for learning rich representations of data, from natural language processing to computer vision. Even in cases that the embedding is not explicit, many algorithms operate over similarity measures which implicitly embed the original dataset. In this thesis, we attempt to understand the intuition behind continuous representations. Can we construct a general theory of continuous representations? Are there general principles for semantically meaninguful representations? In order to answer these questions, we develop a framework for analyzing continuous representations through diffusion limits of random walks. We show that measureable quantities of discrete random walks with a latent metric structure have closed form diffusion limits. These diffusion limits allow us to approximate attributes of the discrete random walk such as the stationary distribution, hitting time, or co-occurrence using closed-form expressions from diffusions. We establish limits which guarantee asymptotic consistency of such estimators, and show they work well in practice. Using this new approach, we solve three classes of problems: first, we derive principled network algorithms which connect statistical estimation tasks such as density estimation to network algorithms such as PageRank. Next, we demonstrate that continuous representations of words are a type of random walk metric estimator with close connections to manifold learning. Finally, we apply our theory to single-cell RNA seq data, and derive a way to learn time-series models without trajectories by using stochastic recurrent neural networks.
Thesis: Ph. D., Massachusetts Institute of Technology, Department of Electrical Engineering and Computer Science, 2016.; Cataloged from PDF version of thesis.; Includes bibliographical references (pages 193-202).
2016-01-01T00:00:00ZOblivious RAM : from theory to practice
http://hdl.handle.net/1721.1/105668
Oblivious RAM : from theory to practice
Fletcher, Christopher W. (Christopher Wardlaw)
Privacy of data storage has long been a central problem in computer security, having direct implications for many Internet-era applications such as storage/computation outsourcing and the Internet of Things (IoT). Yet, the prevailing way we protect our data - through encryption techniques - doesn't protect where we read or write in our data. This additional information, the access pattern, can be used to reverse-engineer proprietary programs as they run, reveal a user's physical location or health information, and more, even if data is correctly encrypted. This thesis studies a cryptographic primitive called Oblivious RAM (ORAM) which provably hides a client's access pattern as seen by untrusted storage. While ORAM is very compelling from a privacy standpoint, it incurs a large performance overhead. In particular, ORAM schemes require the client to continuously shuffle its data in untrusted storage. Early work on ORAM proves that this operation must incur a client-storage bandwidth blowup that is logarithmic in the dataset size, which can translate to > 100x in practice. We address this challenge by developing new tools for constructing ORAMs that allow us to achieve constant bandwidth blowup while requiring only small client storage. A reoccurring theme is to grant untrusted storage the ability to perform untrusted computation on behalf of the client, thereby circumventing lower bound results from prior work. Using these tools, we construct a new ORAM called Ring ORAM, the first small client storage ORAM to achieve constant online bandwidth blowup. At the same time, Ring ORAM matches or improves upon the overall bandwidth of all prior ORAM schemes (given equal client storage), up to constant factors. Next, we more heavily exploit computation at the storage to construct Onion ORAM, the first scheme with constant worst-case and overall bandwidth blowup that does not require heavy weight cryptographic primitives such as fully homomorphic encryption (FHE). Instead, Onion ORAM relies on more efficient additively or somewhat homomorphic encryption schemes. Finally, we demonstrate a working ORAM prototype, built into hardware and taped-out in 32 nm silicon. We have deployed the design as the on-chip memory controller for a 25 core processor. This proves the viability of a single-chip secure processor that can prevent software IP or data theft through a program's access pattern to main memory (having applications to computation outsourcing and IoT). From a technical perspective, this work represents the first ORAM client built into silicon and the first hardware ORAM with small client storage, integrity verification, or encryption units. We propose a number of additional optimizations to improve performance in the hardware setting.
Thesis: Ph. D., Massachusetts Institute of Technology, Department of Electrical Engineering and Computer Science, 2016.; Cataloged from PDF version of thesis.; Includes bibliographical references (pages 117-122).
2016-01-01T00:00:00ZDesigning personal systems for mindful decision making
http://hdl.handle.net/1721.1/105667
Designing personal systems for mindful decision making
Farve, Niaja Nichole
Today's personal technologies are generally seen as reducing mindfulness. Users are so absorbed in their devices that they behave in more distracted ways, are less engaged in face-to-face social interactions and increase their sedentary behaviors. This often results in behaviors and habits that are misaligned with the user's goals. Current attempts to use technology to improve well-being, such as fitness trackers, do not take advantage of some of the benefits that mobile, personal technologies have to offer. Specifically, increasingly mobile personal technologies have the opportunity to intervene in the moment when a person is making a decision with personalized, "just-in- time" nudges that may result in a more mindful decision. This thesis explores how to design personalized, wearable technologies that can support more mindful behavior. It investigates the various challenges that exists when designing such systems-.and provides design considerations for future systems. Human behavior researchers have argued that although a user may have the motivation and the ability to change behavior, a trigger is required to make a new behavior happen. This thesis specifically focuses on considerations that should be made when designing triggers for persuasive, wearable systems. These include ensuring the user's attention, utilizing contextual cues to determine timing of triggers and using personalized messages in a trigger. The thesis presents several pilots studies in using personal, wearable technologies to offer "just-in-time" triggers for behavior. The design and implementation of these systems is detailed and preliminary data regarding their effectiveness is discussed. These systems explore what challenges emerge when applying traditional behavior change theories on personalized, wearable systems.
Thesis: Ph. D., Massachusetts Institute of Technology, Department of Electrical Engineering and Computer Science, 2016.; Cataloged from PDF version of thesis.; Includes bibliographical references (pages 169-177).
2016-01-01T00:00:00Z