It has been a long-standing goal in Artificial Intelligence (AI) to build machines that can solve tasks that humans can. Thanks to the recent rapid progress in data-driven methods, which train agents to solve tasks by learning from massive training data, there have been many successes in applying such learning approaches to handle and even solve a number of extremely challenging tasks, including image classification, language generation, robotics control, and several multi-player games. The key factor for all these data-driven successes is that the trained agents can generalize to test scenarios that are unseen during training. This generalization capability is the foundation for building any practical AI system.

This thesis studies generalization, the fundamental challenge in AI, and proposes solutions to improve the generalization performances of learning agents in a variety of problems. We start by providing a formal formulation of the generalization problem in the context of reinforcement learning and proposing 4 principles within this formulation to guide the design of training techniques for improved generalization. We validate the effectiveness of our proposed principles by considering 4 different domains, from simple to complex, and developing domain-specific techniques following these principles. Particularly, we begin with the simplest domain, i.e., path-finding on graphs (Part I), and then consider visual navigation in a 3D world (Part II) and competition in complex multi-agent games (Part III), and lastly tackle some natural language processing tasks (Part IV). Empirical evidences demonstrate that the proposed principles can generally lead to much improved generalization performances in a wide range of problems.

This paper presents an analysis of Mandarin Chinese as having both singular classifiers and a plural classifier, xie. The Num head is the cross-linguistic locus for number marking, hosting both plural suffixes (such as English -s) and classifiers. As Mandarin bare nouns are general number (inclusive plural), singular classifiers create a singular reading while plural classifier xie creates a plural (exclusive plural) reading; thus Mandarin makes a three-way grammatical number distinction.

Some recent developments in the weakly-coupled heterotic string phenomenology are reviewed. We discuss several important issues such as dilaton/moduli stabilization, supersymmetry breaking (by hidden-sector gaugino condensation), gauge coupling unification (or the Newton's constant), the QCD axion, as well as cosmological problems involving the dilaton/moduli and the axion.

Statistics has a wide application in science and engineering fields. This research work is aim to study the progressive fiber breaking evolution in the longitudinally reinforced composites from a statistical perspective. First of all, the fiber breaking evolution in a single fiber composite is studied. The Kolmogorov-Smirnov goodness-of-fit test is performed on the experimental data to characterize the damage pattern of the fiber in a single fiber composite. The results indicate that the fragmentation evolution of single fiber composites follows the Weibull statistic. Further investigation is focused on the damage initiation sequence in the multi-fiber composites. Four stochastic competing mechanisms are proposed to address the local load redistribution from the broken fibers to the intact fibers. These mechanisms are based on the dominating weakness selection, random walk selection, the self-avoiding walk (SAW) select and the all surrounding neighborhood fiber selection. Finally, the evolutionary fiber breaking process in the multi-fiber composites is demonstrated by using the Weibull statistic to govern the fiber breakings in the longitudinal direction, and the competing models to describe the sequence of damage initiation.

The primary objective of this research work is to investigate the effective mechanical responses of continuous fiber reinforced composites by modifying and extending the available micromechanical framework. A major part of the work conducted involves the investigation of the effective damage responses due to damage evolutions of matrix microcracks and fiber breakages.

Chapter 3 presents the effective elastic damage behavior of continuous fiber reinforced composites with evolutionary matrix microcracks. A cohesive penny-shape microcrack model is proposed within a two-step homogenization framework to achieve the effective elastic damage behavior of continuous fiber reinforced composites. In the proposed model, the size and the number density of microcracks are defined as two damage parameters to control the matrix microcrack evolution. In addition, the thermal effect is taken into account by taking advantage of the thermal eigenstrain and the Eshelby's equivalent inclusion principle. The overall coefficient of thermal expansion (CTE) of the composite is systematically derived under the framework of micromechanics to describe the overall damage behavior of composites due to matrix microcrack evolution under temperature changes.

Chapter 4 proposes a micromechanical evolutionary damage framework capable of predicting the overall mechanical behavior of and damage evolution in continuous fiber reinforced composites. In the framework, the effective stress fields in a single fiber due to an embedded penny-shaped fiber breakage are systematically derived by applying the double-inclusion theory. The notion of effective length denoting the distance between two adjacent breakages is introduced as a damage parameter while determining the damage evolution within a single fiber. This enables the modeling of the effective damage behavior of a single-fiber reinforced composite. As an application of the proposed framework, a micromechanical damage model is further proposed to simulate the fiber-dominated failure mechanism within a multi-fiber composite. A Weibull probability function is adopted to estimate the varying volume fractions of damaged fibers and intact fibers. Numerical simulations are presented to demonstrate the effectiveness of the proposed methodology.

In Chapter 5, based on the linear elastic fracture mechanics (LEFM) and ensemble-volume averaging technique, an effective eigenstrain is newly proposed to quantify the homogenized stress fields in a single fiber due to multiple fiber breakages. In the proposed model, the number density evolution of fiber breakages is characterized by a two-parameter Weibull statistic with the temperature effect implicitly enclosed by properly adjusting the Weibull parameters. The damage criterion in the evolutionary damage model is theoretically derived. Utilization of the proposed damage framework, a homogeneous damage evolution model capable of simulating the material behavior of multi-fiber reinforced composite materials is developed.

Chapter 6 presents two stochastic risk-competing models to simulate the fiber breakage evolution in a multi-fiber composite with an inhomogeneous fashion by considering different load sharing mechanisms. A unit cell model is adopted with each cell being assigned an initial weakness based on a normal distribution. Damage evolution inside each cell structure follows the micromechanical model presented in Chapter 5. Two risk-competing models are introduced subsequently to determine the damage sequence within the multi-fiber composite by computing the fracture probability based on the weakness of cells at each time step. It is observed that one risk-competing model tends to generate a concentrated damage pattern with broken fibers clustering in a T-shape or a cross-shape, while the other model yields a more diffused damage pattern. Finally, the overall stress-strain responses and the fiber breakage evolution are predicted and verified against experimental data.

Chapter 7 examines the effective elastoplastic behavior of metal matrix composites (MMCs) containing unidirectionally aligned continuous fibers. A homogenization procedure is utilized to derive the overall yield function for the composite based on the probabilistic spatial distribution of aligned inclusions. Based on continuum plasticity, a plastic flow rule and a hardening law are postulated. These laws together with the proposed overall yield function then characterized the macroscopic elastoplastic behavior of the composite under three-dimensional arbitrary loading/unloading histories. The overall uniaxial elastoplastic stress-strain behavior of MMCs with aligned continuous fibers is investigated. Comparisons between theoretical predictions and experimental data for the composite are performed to illustrate the capability of the proposed method.

Chapter 8 concludes the present research on micromechanics and effective elastic and elastoplastic behavior of continuous fiber reinforced MMCs. Finally, related future research topics are discussed briefly.

Enabling technologies that transfer micron-sized objects into mammalian cells are needed to advance key applications in cell engineering. High-throughput delivery of organelles, modified intracellular pathogens, enzymes, proteins, and other types of cargo is required to obtain reliable data in settings of cellular heterogeneity. Delivered objects must also avoid entrapment in cell vacuoles or endosomes to retain functionality. Here, we report a massively parallel photothermal microfluidic platform for high-throughput delivery of large cargo directly into the cytosol of mammalian cells. Micron-sized cargo is reproducibly delivered in up to 100,000 cells on the platform in a minute, with parameters for cargo delivery tunable for each cell and cargo type. The delivery platform is a compact silicon chip on which hundreds of thousands of micron-sized cavitation bubbles expand and collapse within nanoseconds in response to pulsed laser excitation. Cavitation bubbles are formed by rapid heat transfer from nearby metallic nanostructures via a photothermal process enhanced by the lightning-rod effect. High speed, localized fluid flows near cavitation bubbles puncture contacting cell membranes with precision, resulting in micron-sized transient membrane pores. Applied pressure provides an active driving force to speed slow diffusing large cargo through these transient pores before membrane repair and resealing. We have reproducibly delivered large cargo including micron-sized bacteria, enzymes, antibodies, and functional nanoparticles into a variety of cell lines, including three different types of primary cells, with high efficiency and high cell viability. Our experiments involving delivery of functional enzymes, antibodies and vacuole escape-incompetent bacteria prove that cargo is delivered directly into the cell cytosol on this platform, bypassing endocytosis. Massively parallel and nearly simultaneous delivery of cargo into cells under the same physiological conditions enables reliable statistical measurements of cargo interactions with cells over time. To demonstrate its utility, we used this platform to explore the intracellular lifestyle of Francisella novicida and discovered that the iglC gene is unexpectedly required for intracellular replication even after vacuolar escape into the cell cytosol. Our photothermal platform approach, termed BLAST (Biophotonic Laser Assisted Surgery Tool), provides a reliable mechanism for high-throughput engineering of cells with micron-sized natural and man-made materials.

Graphene is a carbon based material that has only one atomic layer. It has exceptional electronic and mechanical properties which makes itself an ideal system to study nanoelectromechanical behaviors. In this work I present the fabrication and electromechanical properties of single layer graphene resonator in nanoribbon geometry.

In the past five decades, several landmark discoveries have demonstrated that nucleic acids are a multi-talented class of biomolecules. These discoveries include naturally occurring ribozymes (RNA enzymes), man-made ribozymes and deoxyribozymes (DNA enzymes), and DNA and RNA aptamers (nucleic acid-based molecular binders). These advancements have increased the popularity of nucleic acid molecules as potential biorecognition elements for analytical and biosensor applications. The studies presented here focuses on the in vitro selection of a rational deigned functional aptamer molecules capable of identifying bladder cancer urine samples by recognizing the urinary molecular compositions, or urine biomarkers, associated with bladder cancer. To achieve the aforementioned, I described the different in vitro selection strategies designed to identify aptamer molecules that recognize various target molecules with high specificity and affinity. This is followed by a more detailed focus on the in vitro selection of structure-switching aptamer scheme where I reported the designing and testing of a diverse DNA library. I provided three different selections that were performed utilizing the custom DNA library to demonstrate the feasibility of the library design and the structure-switching in vitro selection platform. Lastly, I performed a complex selection to isolate bladder cancer specific structure-switching aptamers. The complex selection yield specific-aptamers that can distinguish bladder cancer urine samples from hematuria, or blood in urine, urine samples.