Introduction
Texas ScholarWorks was established to provide open, online access to the products of the University's research and scholarship, to preserve these works for future generations, to promote new models of scholarly communication, and to help deepen community understanding of the value of higher education.
UT Tower and campus image credit: Earl McGehee, CC-BY, https://www.flickr.com/photos/ejmc/7452145850
Communities in TSW
Select a community to browse its collections.
Recent Submissions
Enhancing optimization algorithms for non-convex settings : federated and decentralized learning
(2024-08) Tziotis, Isidoros; Mokhtari, Aryan; Shakkottai , Sanjay; Dimakis, Georgios Alexandros; Caramanis, Constantine; Hassani, Hamed
In the last decade, the rise and unprecedented success of neural networks have drawn significant attention from the machine learning community to non-convex optimization. This thesis focuses on federated and decentralized learning within the framework of non-convex optimization, aiming to deepen the understanding of the field and derive novel results. Specifically, the first part of this work addresses two major challenges encountered in the context of Federated Learning: (i) data heterogeneity, where data distribution can vary substantially across clients, and (ii) system heterogeneity, where the computational power of clients can differ significantly. To tackle these challenges, we propose and analyze federated meta-algorithms that leverage ideas from the realm of representation learning. These algorithms incorporate the statistical characteristics of clients' data to adaptively select the clients, thereby speeding up the learning process. Our methods exhibit robust performance in the presence of stragglers, outperforming well-established baselines for both convex and non-convex loss functions. The ability to handle the complexities of federated environments with varying participation structures and resource constraints underscores their practical relevance and potential for real-world deployment. In the second part of this thesis, we delve into Fully Decentralized Learning. We explore the limitations of prior literature and strive to design a novel method that converges to a second-order stationary point in non-convex smooth settings. Our approach involves coupling the gradient tracking scheme with perturbation techniques inspired from the centralized regime. This novel combination allows us to provide, for the first time, polynomial convergence rate guarantees with respect to the last iterate. Our experimental validation demonstrates that our proposed method efficiently escapes saddle points, achieving superior performance compared to existing decentralized algorithms.
The charisma conundrum
(2024-08) Davis, Emily Anne ; Pangle, Lorraine Smith; Stauffer, Devin; Weyland, Kurt G; Suri, Jeremi; Gilmore, Nathaniel K
In this dissertation, I examine the concept of charismatic leadership through works by four political theorists: Weber, Xenophon, Plutarch, and Stendhal. Despite its political salience, charismatic leadership remains ill-defined. A major reason for this is that Weber, the first thinker to identify it as a concept, does not clearly explain what drives the charismatic leader or what his people hope to receive from him. I argue that Weber leaves these issues ambiguous because he treats charismatic leadership in the context of a larger framework of human motivation, which separates moral incentives from self-interested ones. Weber’s treatment raises the possibility that these incentives combine to motivate the charismatic leader and his people, but his framework prevents him from exploring this possibility. I turn to ancient texts—specifically, Xenophon’s Education of Cyrus and Plutarch’s Life of Caesar—for further clarification, as the influence of ancient thought and history on Weber’s work suggests that the phenomenon of charismatic leadership has ancient roots. And, indeed, Xenophon’s and Plutarch’s portraits of Cyrus and Caesar mirror Weber’s portrait of the charismatic leader. But their portraits illuminate the motivations that Weber’s leaves obscure. Xenophon and Plutarch show that charisma consists in the leader’s ability to heighten his people’s hope for a perfectly just world: one in which they can not only lose themselves in service to a noble cause, but also be consistently rewarded for their virtue. The leader makes this impact because he, too, longs for such a world, so much so that he believes he can create it. Yet, enchanted by visions of his own heroism, he begins to crave unparalleled gratitude and honor, which drives him to amass supreme power at his people’s expense. The same phenomenon, I demonstrate, appears in Stendhal’s writings on Napoleon, which, along with Weber’s references to ancient thought, reveal a through-line from the ancients to the moderns on charismatic leadership. My dissertation, by providing a clearer picture of charismatic leadership than previous literature has, evinces the usefulness of comparing ancient and modern thought on this important topic.
Chemically enabled CO₂-enhanced oil recovery
(2024-08) Pang, Jieqiong ; Mohanty, Kishore Kumar; Chun Huh; Keith P. Johnston; Ryosuke Okuno; Matthew Balhoff
Carbon-dioxide enhanced oil recovery (CO₂-EOR) is widely used for oil recovery from reservoirs. Additionally, the interest in CO₂ storage in depleted reservoirs and aquifers grows due to global warming. However, continuous gas injection suffers from low sweep efficiency caused by unfavorable mobility ratio, gravity segregation, and reservoir heterogeneity. The water-alternating-gas (WAG) was proposed to overcome these challenges, but still encounters issues. Chemically enabled CO₂-EOR methods, involving the adding of surfactants, polymers, and nanoparticles, are explored to address these challenges. This dissertation aims to develop effective CO₂-EOR strategies, including CO₂ polymer-alternating-gas (CO₂-PAG) and CO₂-foam flooding, through a series of experiments and simulations. Minimum miscibility pressure (MMP) is an important parameter in gas injection projects, indicating the pressure at which injected gas and reservoir oil become miscible during flow. CO₂ floods are efficient at or above MMP, enhancing oil recovery. Slim tube experiments were conducted to measure MMP for various oils and the effects of CO₂ additives and rich gas content on MMP were studied. Conclusions include MMP values for different oils and the limited impact of certain additives on miscibility. Compositional simulations suggested rich gas can notably reduce MMP between crude oil and CO₂. Next, the performance of CO₂ polymer-alternating-gas (CO₂-PAG) flooding for enhanced oil recovery was evaluated through core flood experiments and 2-D visualization experiments, followed by simulation studies. Core flood experiments were conducted using both homogeneous and heterogeneous core samples, with a novel heterogeneous core system developed to mimic layered reservoirs. Incremental oil recovery from PAG flooding was compared to water-alternating-gas (WAG) floods and continuous gas (CG) floods. Additionally, two-dimensional (2-D) visualization flooding experiments were performed in layered heterogeneous sand-packs. Subsequent simulations were conducted to further analyze PAG performance and refine strategies. Findings reveal that PAG flooding demonstrated improved performance in heterogeneous systems, compared to CG and WAG floods. Subsequently, the efficiency of CO₂-foam flooding with wettability alteration was investigated in an oil-wet carbonate reservoir. Three types of surfactants, including a nonionic surfactant Soloterra 843, a cationic surfactant CETAC-30, and a zwitterionic surfactant LMDO, were evaluated as potential foam agents. Contact angle tests using core chips were conducted to assess the wettability alteration capability of these surfactants. Subsequently, foam static stability tests and foam rheology tests were conducted under reservoir conditions to evaluate the performance of various foam formulations. Finally, foam core flood experiments were performed using carbonate core samples. Results from contact angle experiments indicated that CETAC-30 could effectively alter rock wettability from oil-wet to water-wet. Foam stability tests revealed that LMDO exhibited superior stability in generating CO₂-foam in the absence of crude oil, followed by CETAC-30 and Soloterra 843. However, the presence of crude oil adversely affected CO₂-foam stability. Rheological tests demonstrated that the apparent viscosity of CO₂-foam decreased with increasing flow velocity, exhibiting a shear thinning behavior. Furthermore, core flood experiments demonstrated that injecting wettability alteration surfactant, CETAC-30, in form of foam could not only alter the wettability of rocks but also enhance the foam strength. However, the introduction of pre-flushed wettability alteration surfactant may not significantly impact foam strength. Additionally, most of the wettability alteration surfactant injected with CO₂ tended to be used in foam generation, with any remaining surfactant utilized for wettability alteration. Finally, CO₂-foam flooding was evaluated as a method to improve CO₂ storage in a high salinity carbonate reservoir. Several surfactants and nanoparticles were examined to identify an effective foam formulation. Foam stability at the ambient pressure and at the reservoir pressure was used to screen suitable foaming agents. The foam mobility was measured through a carbonate rock at 80% quality with selected foaming agents. Finally, CO₂ flooding, and CO₂-foam flooding experiments were performed in carbonate core samples under the reservoir conditions. Many surfactants and nanoparticles were insoluble in brine at the high salinity and temperature conditions of this study. Through foam stability and rheology tests at the reservoir pressure, a combination of the nonionic surfactant Aspiro S 2410 and the nanoparticle EOR 12-V3 was found to be the most effective. The addition of nanoparticles significantly increased the half-life of the foam at the reservoir pressure. The presence of crude oil had a detrimental effect on CO₂-foam stability. The surfactant-nanoparticle foam exhibited a higher apparent viscosity compared to the foam generated by the surfactant alone. Core flood experiments demonstrated that CO₂-foam flooding with the surfactant-nanoparticle solution achieved higher CO₂ storage and oil recovery compared to both continuous CO₂ injection and CO₂ surfactant-foam flooding.
Development of the ACT hand with a broad and precisely adjustable stiffness spectrum via series elastic actuators
(2024-08) Bae, Jung Hyun ; Deshpande, Ashish D.; Raul Longoria
Tendon-driven systems offer significant advantages for robotic manipulators, including low inertia and improved force control. The Anatomically Correct Testbed (ACT) robotic hand is one such system, developed to understand the intrinsic biomechanical and control features of human hands. However, this system has limited compliance, and controlling it is challenging due to its inherent nonlinearities and nonholonomic constraints. In this research, we integrated Series Elastic Actuators (SEAs) into the ACT Hand to address these challenges. SEAs were chosen for their ability to improve compliance by reducing environmental impact forces and significantly expanding the stiffness range spectrum, enabling easier achievement of the desired stiffness while maintaining stability. Unlike the original ACT Hand, where stiffness was passively determined, the SEA configuration allows the stiffness of the robot finger's end-tip in Cartesian space to be actively adjusted through both the physical manipulation of each SEA's stiffness and the positional control of the motors connected to the tendons. In the ACT Hand structure, the six tendons of the index finger, each serially connected to a spring for compliance, actuate four degrees of joint freedom to achieve the desired end-tip position. By implementing SEAs, we aimed to enhance adaptability and enable precise adjustment of the end-tip stiffness in the ACT Hand. To validate the effect of SEAs on adjustable stiffness, we conducted experiments by varying the stiffness of the SEAs and altering tendon tension via positional control. We then compared the stiffness ellipses of the ACT Hand index finger's end-tip with and without SEA implementation. The comparison revealed that SEAs enhanced the precision and control of adjustable stiffness across a broad spectrum. This improvement demonstrates that the implementation of SEAs in the ACT Hand enables the passivity bound for conservative stability to be easily maintained through active stiffness variation via position control. Additionally, SEAs enhance the ACT Hand's capability to adapt to various environments by fine-tuning stiffness in each direction within the Cartesian space.
Securing access control using machine learning and formal methods
(2024-08) Hu, Yang, Ph. D. ; Tiwari, Mohit; Vijay K. Garg; Christine Julien; Kenneth McMillan; Sarfraz Khurshid
Access control is a fundamental security mechanism for computer systems, acting as the first line of defense against potential threats. Its primary objective is to prevent unauthorized access that violates the access control policies/configurations predefined by system administrators. However, access control vulnerabilities and misconfiguration can introduce additional attack surfaces to the systems. This can lead to severe security breaches and financial losses. To enhance access control security, we introduce three novel techniques that leverage machine learning and formal methods to detect or fix access control vulnerabilities or misconfigurations in operating systems and cloud systems. First, we present ACHYB, a hybrid program analysis technique to detect access control vulnerabilities in the Linux Kernel. Second, we introduce TAC, a greybox verification based penetration testing technique which utilizes deep reinforcement learning and IAM modeling to detect privilege escalations in cloud access control misconfigurations. Third, we propose IAMPERE, which utilizes Graph Neural Networks and MaxSAT to automatically fix privilege escalations in cloud access control misconfigurations. Evaluation on synthetic and real-world benchmarks demonstrates that all three techniques show promising effectiveness and efficiency compared to existing methods.