116964 research outputs found
Sort by
Improving online decision making with prior data
No full textSystems that learn to make optimal decisions with online interactions with an unknown environment are commonplace in the modern world. These include recommendation engines for online marketplaces, automated diagnosis models in healthcare and autonomous vehicular controllers among several others. The quality of their decisions is crucial to enhancing the experience of users that employ them. Due to the high cost or negative impacts associated with learning optimal policies from scratch, these systems are usually trained with a combination of offline and online data, with the former being used to arrive at a reasonable first-pass policy that is then fine-tuned with online interaction. While the offline data is generally abundant, it is often the case that it was collected under different environmental parameters, which can introduce significant biases in the pre-training phase leading to the learning of suboptimal policies online. Quantifying the shifts in the environments between the two phases or identifying invariant statistics in the two, however, can lead to significant improvements in the speed of learning optimal policies online. In this thesis, we consider three problems in this space and provide solutions for each of them that improve over naive methods that do not use the prior data. First, we consider the problem of Episodic Bandits with Stochastic Experts where the agent interacts with a the environment through a set of experts over episodes. With full knowledge of expert policies and context distributions, we propose a strategy that leverages the information leakage between experts and guarantees a regret upper bound that does not scale in the horizon of interaction. We extend this the case when experts are only known approximately and context distributions are not revealed. We show that with sufficient precision in the empirical estimates, one can recover the constant regret guarantee, similar to the full-information setting. Finally, we show that these estimates of expert policies can be inferred using historical interactions and used to bootstrap the online phase. Next, we study the problem of Bandits with Mean Bounds where an agent is provided with upper and lower bounds on the mean rewards of arms that it must choose from. We study both the stochastic multi-armed setting as well as a linear case and propose novel strategies that use the side information in order to restrict the set of arms to be considered and modify their exploration rates. These lead not only to improvements in regret performance but also to significantly reduced computation. We show that such mean bounds occur naturally when one has access to confounded prior data. Here, confounding means that the policy used to collect the data observed more information than was recorded. We provide novel schemes to extract mean bounds from confounded logs and use them in conjunction with our online algorithms to improve regret performance. Finally, we treat the problem of Hot Object Cache admission policy selection in Content Delivery Networks. Traditionally, these policies have been parameterized by static thresholds on request size and frequency. However, with changing input traffic, the best performing parameters also vary. Thus, there is a need for adaptive policies that modify the parameters according to the incoming traffic statistics. We propose a learning framework that consists of three stages: feature estimation, cluster selection and neural bandit expert selection. The first stage forms an approximate view of the environment through visible statistics and the second uses offline trained clusters to reduce the space of policies to be considered. Finally, the third phase models the expert selection as a Best Arm Identification with Side Information problem. We provide a policy and show analytically that the mean time to terminate does not scale with the number of experts. We also present empirical simulations that show that our proposed method significantly outperforms all static policies as well as existing learned baselines.Electrical and Computer Engineerin
Aspects of SU(4) ferromagnetism in strongly correlated graphene
Full text linkThis dissertation presents studies of correlated phases of matter and their transport properties in graphene-based correlated systems. A common thread is the spontaneous breaking of the approximate SU(4) spin-valley symmetries. Chapter 1 is a brief introduction to three types of strongly correlated graphene studied in the thesis. Chapter 2 focuses on the monolayer graphene in the strong magnetic field and the collective modes such as spin waves of the integer quantum Hall states that spontaneously break the SU(4) spin-valley symmetries in N = 0 Landau level. A systematic numerical method is developed to calculate the scattering of collective modes by gate-controlled junctions between quantum Hall states with different integer filling factors ν. It is found that the dispersion mismatch suppresses the transmission probability of collective mode, shedding light on related experiments. Chapter 3 proposes a domain wall depinning mechanism for the current-induced magnetization and Hall conductance reversal in the Chern insulators arising from spontaneous valley (orbital) polarization, in particular, the quantum anomalous Hall state discovered in the magic angle twisted bilayer graphene. Chapter 4 analyzes a possibility to detect time reversal invariant intervalley coherent states from the transport measurement. Our results show that the intervalley coherent orders can in principle give rise to the weak (anti-)localization effect in graphene and can be identified in the low-temperature weak-field magnetoresistance measurements. In the end, Chapter 5 discusses several open questions and interesting future directions related to the strong correlation physics in graphene and other two-dimensional material platforms.Physic
Improving human machine interaction by striking a balance between device control and biosignal quality for users with various levels of ability
Full text linkBiosignals have recently gained attraction as a tool to control human machine interaction (HMI) for applications ranging from assistive device control to entertainment. A successful HMI system requires an intuitive way for the user to send their commands, an accurate method to recognize user's intention, and a responsive device control method to help the user achieve their goal. Electromyography (EMG), a technique for recording muscle activation as electrical signals, is an intuitive tool for the users to send their commands. Extracting user's intention from noisy EMG signals is a non-trivial problem. Moreover, users with disabilities have lower-quality biosignals and less control over their muscle activation. Depending on the quality and availability of biosignals from the user, information such as opening or closing of the hand, limb position, or force could be extracted from EMG signals. As a final piece of the HMI system, for each application, a safe responsive device control method needs to be developed that provides the user with maximal functionality and does not hinder their ability during daily tasks. My overarching idea is that there is a trade-off between the achievable HMI capabilities and the control complexity of the HMI. And by matching the complexity of HMI control algorithm to the biosignal quality and needs of the target population, capabilities of the HMI system could be optimized. To this end, I present bio-signal based device control methodologies to address unique needs of users with various levels of hand ability and I evaluate the HMI performance. For users with severe hand disability, I present a methodology for EMG-based intention recognition and assistive hand exoskeleton control that ensures accurate, safe and robust control during daily activities and prevents fatigue by reducing neuromuscular effort. Through experimentation with able-bodied and affected subjects, I demonstrate that the developed method robustly classifies user's intention and assists them while reducing their neuromuscular effort. For users with mild hand disability, I present novel intention recognition and exoskeleton control methodologies to maximize functionality of the assistive device by allowing the user to simultaneously control the grasping pose as well as grasping strength. This method augments user's grasping strength proportional to their intended force, yet keeps them engaged in performing the task. Through pilot studies, I confirm user's success to activate the intended hand pose and accurately follow dynamic force trajectories using the assistive HMI system. Finally, for able-bodied users, I propose data-efficient EMG-based method for dynamic task detection which can be used as an input to HMI. I develop and optimize two similarity-based methods for biosignal analysis based on Euclidean distance and correlation, then I compare their performance against state-of-the-art binary classification methods. I demonstrate that the developed methods can accurately detect the dynamic finger task and avoid misclassification during other daily tasks, while using very few training data. Overall, I developed bio-signal based HMI solutions for unique needs of three groups of target users with different levels of hand ability, and demonstrated that striking a balance between control complexity and biosignal quality could improve HMI performance compared to state of the art. The methodologies and findings from this research pave the way for accurate, robust, and efficient HMI system development that can optimize the functionality of HMI systems such as assistive exoskeletons and allow the user to maximize their capability while avoiding fatigue, unsafe interactions, disengagement, or misidentification of intent. These principles would unlock opportunities to develop HMI systems that are even further customizable to users' abilities and preferences and can respond to users' state or changing needs.Mechanical Engineerin
Impact of Legacy Wells on Carbon Storage and Energy Development in Texas: Cost Assessment, Policy Insights, and Management Strategies. A Technical White Paper
Full text linkTexas has a unique opportunity to lead in clean energy innovation (carbon storage, hydrogen
storage, and others) by reusing the same subsurface formations that have been utilized by the
oil and gas industry for decades. This opportunity comes with a challenge: Texas has hundreds
of thousands of legacy oil and gas wells, many drilled before 1970 and lacking construction and
plugging records that are needed to ensure that wells have zonal isolation (non-leakage). When
new injection projects increase underground pressure, these non-isolating wells can allow fluids
to move from the injection zone to shallower zones where they could leak to damage freshwater,
the surface or the atmosphere.
Experience with leaking wells shows that risks vary by location. In the Permian Basin, injection
of large volumes of produced water has raised underground pressure and, in some places,
caused unmanaged wells to leak brine. In contrast, the Gulf Coast’s younger, more permeable
formations dissipate pressure, reducing the risk of interference even under high injection vol-
umes. Building on these statewide findings, this study focuses on two typical high-well density
counties: Crane County (Permian Basin) and Chambers County (Gulf Coast) to explore how
legacy wells shape the risk and cost structure of future clean energy projects, and to show the
value of well plugging programs for resource development. Our pilot study of two representa-
tive prospective CO2 storage areas of Texas show how geology and well density influence the
cost of preparing a site to ensure safe storage. We estimate the costs of preparing legacy wells
for injection projects for each county based on geologic properties controlling fluid acceptance
and the distribution of wells that must be prepared.
The estimated mean cost of evaluating and mitigating all wells within the area review of a typical 1.5 to 2.8 million metric tons (MMT) carbon storage injection project into the San Andres
Formation in Crane County around 5.6 million. However, wells are strongly clustered in producing fields with very high density areas giving the highest 10-20 million costs, and 70% of
storage resource areas have legacy well preparation costs less than the average. In Chambers
County, the nominal project size is increased to 20 MMT injection over 20 years, because the
Miocene injection zone is thicker, more compressible, and has higher porosity and permeability
than the San Andres and can accept these volumes. Despite the larger injection volumes, the
Chambers County area of reviews are consistently smaller, and therefore average well mitiga-
tion costs are around 1.5 million per standard carbon storage project. Wells in the Gulf Coast
are even more strongly clustered in producing fields, so that 79% of the county has legacy well
preparation costs less than average.
The data collected in this study illuminates the cost of legacy well management for new project
development. The cost is not evenly distributed; some areas have minimal legacy wells costs
but in others the cost is likely to be prohibitively high. Resource recovery projects that require
well management may be exceptions. The variable well remediation cost may have an impact
on acreage leasing value. Consideration of well density may be relevant to state investment in
well remediation either through well programs.Bureau of Economic Geolog
Philanthropy in higher education : making meaning of development officer social identity during donor cultivation
Full text linkThe purpose of this study was to make meaning of the experiences of higher education development professionals, during donor engagement activity, and understand how social identity impacts giving. Development officer experiences were captured via semi-structured interviews. Two research questions were used: 1) how do development officers at a public four-year institution of higher education (IHE) make meaning of experiences they have when cultivating donors? and, 2) how do development officers at a public four-year IHE perceive the impact of their social identity (race/ethnicity, gender, religion, sexual orientation, physical appearance, age) during donor engagement? Social cognitive theory, which utilizes environmental and behavioral cues, was used to frame the responses of the development officers. Planned behavior theory was used to connect development officer behavior to beliefs, and to interpret gender interaction, racial perceptions, sexual orientation interaction, and physical appearance prejudices. Two underlying questions within this study were, 1) will the goal of obtaining a gift outweigh a development officer’s personal discomfort with a donor’s language, behavior, or opinions? and 2) is the donor’s desire or decision to make a gift impacted when they are placed into an environment, or interact with social identities, which influences them (as observed by the development officer). This study assumed: 1) all of the experiences of the development officers are unique, 2) development officers understand and recognize social identity factors which may influence giving, 3) development officers are able to make a connection between their social identity and the influence it may have on donor giving, and. 4) development officers are able to honestly reflect upon, and properly convey, their donor engagement experiences. This study also assumed that social identity influence on donor giving exists but may not occur in every interaction, and that donors would not willingly convey negative reasons why they were influenced (i.e. sexism, racism, religious, ageism, prejudice) and might not convey positive reasons either. However, it was also assumed that development officers would be cognizant of all these reasons and be willing to share their beliefs.Educational Leadership and Polic
Crocodylus acutus (American Crocodile). Diet.
No full textThis video demonstrates behavior. Most videos in this collection have no audible language and for those that do, the language isn't necessary to understand the behavior. For that reason, transcripts are not provided.Integrative Biolog
Secluded capital : Baizabai Shinde and the transnational opium trade in nineteenth century South Asia
Full text link“Secluded Capital” is a history of capitalism and gender based on the career of the dowager queen of Gwalior, Baizabai Shinde. She was a successful banker and an influential politician in the nineteenth century South Asia who restricted the colonial ambition of establishing control over the opium trade based in central India (Malwa opium trade). After the British East India Company established its political control over eastern India in the mid-eighteenth century, it gradually promoted trade with China in Bengal opium. The opium trade was critical for the Company as its profits subsidized the Company’s administrative apparatus in India. By the early nineteenth century, the Company had monopolized the production and sale of opium in eastern India. As the profits of this trade stabilized, the Company discovered that the west coast based Malwa opium trade was undercutting its profits. Despite several attempts, the Company failed to counter the competing Malwa opium trade. The Malwa opium trade was controlled by women-centric network of indigenous states, bankers, and traders. During the peak of the opium trade in the nineteenth century, Baizabai Shinde was the leader of the anti-colonial indigenous economic network. I argue that women-centric aristocratic Maratha customs and practices in political diplomacy, alliance building, and religious patronage pioneered by aristocratic Maratha women were instrumental for Baizabai’s emergence as the unchallenged leader of the indigenous transnational opium trade network. In the chapters of the dissertation, I explore these themes in depth. Baizabai exemplifies the broader trend of the prominence of aristocratic women in the Maratha states as adroit leaders and bankers. The Marathas practiced caste-specific endogamy resting on cross-cousin marriages. From the kin and affines derived from this endogamy, Baizabai built clan-based resources for establishing control over the opium-producing areas of central India and resisting hostile British policies towards Gwalior by ensuring the Shinde clan’s dynastic perpetuation. She expanded this network for including non-kin indigenous political allies (Rajputs) through the creative deployment of ‘fictitious-kinship,’ which ensured her control over opium transmission routes. As Baizabai dictated these policies of alliance-building, in Gwalior, the zenana emerged as the power-center of the Malwa opium trade. Through Baizabai’s example, I demonstrate that the seemingly oppressive elite practices, such as parda, were perceived by aristocratic women as a resource for competing with male colonial officers. Finally, I show how the Maratha noble women such as Baizabai, reinvented themselves as generous religious patrons and encouraged pilgrimage to sacred Hindu cities such as Banaras. This effort not only popularized their rule but also promoted banking through the pilgrims’ demand for credit. Because of the efforts of the Maratha queens, Banaras emerged as the most important sacred city and banking center in northern India. For accomplishing this work, I have consulted multilingual vernacular sources from various archives in India and UK. Particularly, I have used Marathi records in Modi script. I am the first scholar to access the vernacular record of the Gwalior state.Histor
Raw materials for the construction of a city : ceramic production and consumption in the configuration of communities of practice in colonial Panama
Full text linkMy dissertation examines the early colonial history of Panama through the study of production and consumption of ceramics. It echoes postcolonial debates that introduced notions of agency in the configuration of colonial experiences. Previous archaeological research in Panama emphasizes the disappearance of indigenous ceramic styles and their replacement by European and creole styles following the Spanish Conquest. Scholars have interpreted stylistic discontinuity as evidence of indigenous extinction or assimilation in areas the Spaniards occupied. Contrary to this approach, I show the continuity of indigenous technological traditions and their participation in the construction of colonial Panama, by identifying the raw materials and technologies used in ceramic production during the pre-Columbian and colonial periods. This study focuses on compositional and technological analyses including neutron activation analysis (NAA) and petrographic thin section analysis of the ceramic body (pastes) of 250 samples from the pre-Columbian and colonial periods (AD 700-1800) collected at indigenous and Spanish colonial settlements. I assess this evidence using the “communities of practice” concept to frame the dynamic nature of interaction and knowledge transmission around crafting activities. This theoretical approach permits the study of production and consumption of ceramics and the transformation, emergence, and configuration of communities of practice after European colonization. The findings from this research have strong implications for the revision of colonial Panamanian history. Historical narratives about indigenous people in Latin America affect contemporaneous notions of indigeneity and, very often, lead to the exclusion of these groups from processes of nation-building. Based on the importance of these representations, my research contributes to the creation of new versions of history that emphasize the participation of non-European actors and their contribution to the colonial order. Apart from diversifying historical Panamanian narratives, my research contributes to closing the gap between precolonial and colonial archaeologies where indigenous histories get lost.Anthropolog
Unruly voices : narration of communal memory and the construction of gender and communal identity in Assia Djebar’s Far from Madina
Full text linkAssia Djebar’s Far from Madina retells the stories of the women who appear on the margins of the earliest sources of Islamic history from a contemporary Muslim feminist’s perspective. Djebar uses formal elements of early Islamic historiography and relies upon classical Sunni sources. These techniques place her novel in conversation with classical Islamic tradition and bring legitimacy to her subversive project which aims to shift the boundaries of that canon. Though crafted in relation to classical sources, Djebar’s critique of gender identity is also addressed to the discourses and institutions of Islamic authority that evolved over the centuries and that continue to delineate narrow roles for women, up to and including contemporary regimes. In chapter one I argue that by grounding her critique of circulating discourses on Muslim women within a project that appropriates canonical Sunni historiography, Djebar refuses the disjunction between feminism and Islam, critiquing normative Islamic discourse on women in contemporary Algeria without framing the conflict in terms of an East/West or a religious/secular binary. Chapter two examines Djebar’s treatment of Fatima in particular. I consider Djebar’s selection of classical sources and compare the earliest canonical Sunni renderings of Fatima and those found in the novel. I argue that the vision of empowered women in the first Muslim community posited in Far from Madina destabilizes the ideal of gender identity constructed in early Islamic historiography. Far from Madina focuses on the moment after the death of Muhammad when Muslims were left to interpret their scripture and recall their Prophet’s words and deeds. Djebar constructs the novel around the question of what role Muslim women would play in this process, a move which foregrounds her own choice to write the novel and embrace her role as witness and transmitter of the stories of these early women. Chapter three examines the reflexive character of Far from Madina and considers how Djebar’s narrative strategies and hermeneutical approach facilitate the articulation of identity through difference. I argue that the narrative is Djebar’s performance of contemporary Muslim identity and an example of “lived Islam.
Electronic structure and optical properties of B-III-V compound
Full text linkHighly mismatched semiconductor alloys offer unique combinations of bandgap and lattice constant, making them attractive for a wide range of applications. Alloying boron pnictides, which have relatively small lattice constants, into conventional III-V semiconductors presents a promising approach for developing near-infrared, direct bandgap materials that can be lattice-matched to silicon or GaAs. However, B-III-V alloys remain underexplored, and there are conflicting reports on their electronic and optical properties, making it challenging to benchmark and optimize the design and growth of B-III-V materials and devices.
In this work, we employed density functional theory (DFT) with HSE06 hybrid functionals to study the intrinsic mechanical, electronic, and optical properties of BGa(In)As. The theoretical findings were compared with the properties of BGa(In)As grown via molecular beam epitaxy (MBE), as well as extending the analysis to include comparisons with other highly mismatched alloys. We demonstrated the distinct impact of boron incorporation, distinguishing its effects from those observed in other mismatched systems. The study of the electronic structure and intrinsic optical properties of B-containing compounds at various B concentrations revealed the potential of B-III-V alloys for device applications in targeted wavelength ranges, providing valuable insights for the design of B-III-V optoelectronic devices.
Additionally, we investigated the behavior of B-III-V alloys during post-growth annealing, highlighting the mechanisms and benefits of annealing in optimizing the growth and performance
7
of highly mismatched alloys. Our findings provide insights that can guide the design and development of B-III-V compounds for optoelectronic applications.Electrical and Computer Engineerin