Shotgun Proteomics is the leading technique for protein identification in complexmixtures. However, it produces a large amount of data which results in aextremely high computational cost for identifying the protein. Retention time(RT) is an important factor to be used to enhance the efficiency of protein identification.By predicting the retention time successfully, we could decrease thecomputational cost dramatically. This thesis uses a machine learning method,Gaussian Processes, to predict the retention time of a set of peptide in hand.We also implement a feature extraction method called Bag-of-Words to generatethe features for training the model. In addition, we also investigate theeffect of different types of optimization methods to the model’s parameters.The results show comparable precision of the prediction and relatively lowtime cost when comparing with the state-of-art prediction model.

A research report submitted in partial fulfillment of the requirements for the degree of Master of Science in School of Computer Science and Applied Mathematics to the Faculty of Science University of Witwatersrand, 2019 / Learning-based methods have recently become popular in control engineering, achieving good performance on a number of challenging tasks. However, in complex environments where data efficiency and safety are critical, current methods remain unsatisfactory. As a step toward addressing these shortcomings, we propose a learning-based approach that combines Gaussian process regression with model predictive control. Using sparse spectrum Gaussian processes, we extend previous work by learning a model of the dynamics incrementally from a stream ofsensory data. Utilizinglearned dynamics and model uncertainty, we develop a controller that can learn and plan in real-time under non-linear constraints. We test our approach on pendulum and cartpole swing up problems and demonstrate the benefits of learning on a challenging autonomous racing task. Additionally, we show that learned dynamics models can be transferred to new tasks without any additional training. / TL (2020)

No description available.

No description available.

No description available.

We study the problem of Reinforcement Learning (RL) for Unmanned Aerial Vehicle (UAV) navigation with the smallest number of real world samples possible. This work is motivated by applications of learning autonomous navigation for aerial robots in structural inspec- tion. A naive RL implementation suffers from curse of dimensionality in large continuous state spaces. Gaussian Processes (GPs) exploit the spatial correlation to approximate state- action transition dynamics or value function in large state spaces. By incorporating GPs in naive Q-learning we achieve better performance in smaller number of samples. The evalua- tion is performed using simulations with an aerial robot. We also present a Multi-Fidelity Reinforcement Learning (MFRL) algorithm that leverages Gaussian Processes to learn the optimal policy in a real world environment leveraging samples gathered from a lower fidelity simulator. In MFRL, an agent uses multiple simulators of the real environment to perform actions. With multiple levels of fidelity in a simulator chain, the number of samples used in successively higher simulators can be reduced. / Master of Science

In this dissertation, we propose decentralized and scalable algorithms for Gaussian process (GP) training and prediction in multi-agent systems. The first challenge is to compute a spatial field that represents underwater acoustic communication performance from a set of measurements. We compare kriging to cokriging with vehicle range as a secondary variable using a simple approximate linear-log model of the communication performance. Next, we propose a model-based learning methodology for the prediction of underwater acoustic performance using a realistic propagation model. The methodology consists of two steps: i) estimation of the covariance matrix by evaluating candidate functions with estimated parameters; and ii) prediction of communication performance. Covariance estimation is addressed with a multi-stage iterative training method that produces unbiased and robust results with nested models. The efficiency of the framework is validated with simulations and experimental data from field trials. The second challenge is to perform predictions at unvisited locations with a team of agents and limited inter-agent information exchange. To decentralize the implementation of GP training, we employ the alternating direction method of multipliers (ADMM). A closed-form solution of the decentralized proximal ADMM is provided for the case of GP hyper-parameter training with maximum likelihood estimation. Multiple aggregation techniques for GP prediction are decentralized with the use of iterative and consensus methods. In addition, we propose a covariance-based nearest neighbor selection strategy that enables a subset of agents to perform predictions. Empirical evaluations illustrate the efficiency of the proposed methods / Doctor of Philosophy / In this dissertation, we propose decentralized and scalable algorithms for collaborative multiagent learning. Mobile robots, such as autonomous underwater vehicles (AUVs), can use predictions of communication performance to anticipate where they are likely to be connected to the communication network. The first challenge is to predict the acoustic communication performance of AUVs from a set of measurements. We compare two methodologies using a simple model of communication performance. Next, we propose a model-based learning methodology for the prediction of underwater acoustic performance using a realistic model. The methodology first estimates the covariance matrix and then predicts the communication performance. The efficiency of the framework is validated with simulations and experimental data from field trials. The second challenge regards the efficient execution of Gaussian processes using multiple agents and communicating as little as possible. We propose decentralized algorithms that facilitate local computations at the expense of inter-agent communications. Moreover, we propose a nearest neighbor selection strategy that enables a subset of agents to participate in the prediction. Illustrative examples with real world data are provided to validate the efficiency of the algorithms.

Manufacturing systems are usually nonlinear, nonstationary, highly corrupted with outliers, and oftentimes constrained by physical laws. Modeling and approximation of their underly- ing response surface functions are extremely challenging. Bayesian optimization is a great statistical tool, based on Bayes rule, used to optimize and model these expensive-to-evaluate functions. Bayesian optimization comprises of two important components namely, a sur- rogate model often the Gaussian process and an acquisition function often the expected improvement. The Gaussian process, known for its outstanding modeling and uncertainty quantification capabilities, is used to represent the underlying response surface function, while the expected improvement is used to select the next point to be evaluated by trading- off exploitation and exploration. Although Bayesian optimization has been extensively used in optimizing unknown and expensive-to-evaluate functions and in hyperparameter tuning of deep learning models, mod- eling highly outlier-corrupted, nonstationary, and stress-induced response surface functions hinder the use of conventional Bayesian optimization models in manufacturing systems. To overcome these limitations, we propose a series of systematic methodologies to improve Bayesian optimization for engineering design and quality control of manufacturing systems. Specifically, the contributions of this dissertation can be summarized as follows. 1. A novel asymmetric robust kernel function, called AEN-RBF, is proposed to model highly outlier-corrupted functions. Two new hyperparameters are introduced to im- prove the flexibility and robustness of the Gaussian process model. 2. A nonstationary surrogate model that utilizes deep multi-layer Gaussian processes, called MGP-CBO, is developed to improve the modeling of complex anisotropic con- strained nonstationary functions. 3. A Stress-Aware Optimal Actuator Placement framework is designed to model and op- timize stress-induced nonlinear constrained functions. Through extensive evaluations, the proposed methodologies have shown outstanding and significant improvements when compared to state-of-the-art models. Although these pro- posed methodologies have been applied to certain manufacturing systems, they can be easily adapted to other broad ranges of problems. / Doctor of Philosophy / Modeling advanced manufacturing systems, such as engineering design and quality moni- toring and control, is extremely challenging. The underlying response surface functions of these manufacturing systems are often nonlinear, nonstationary, and expensive-to-evaluate. Bayesian optimization, a statistical modeling approach based on Bayes rule, is used to rep- resent and model those complex (i.e., black-box) objective functions. A Bayesian optimiza- tion model consists of a surrogate model, often the Gaussian process, and an acquisition function, often the expected improvement. Conventional Bayesian optimization models do not accurately represent non-stationary and outlier-corrupted functions. To overcome these limitations, we propose a new asymmetric robust kernel function to improve the model- ing capabilities of the Gaussian process model in process quality control through improved defect detection and classification. We also propose a non-stationary surrogate model to improve the performance of Bayesian optimization in aerospace process design problems. Finally, we develop a new optimization framework that models and optimizes stress-induced constrained aerospace manufacturing systems correctly. Our extensive experiments show significant improvements of these three proposed models when compared to state-of-the-art methodologies.

The design of unconventional aircraft requires early use of high-fidelity physics-based tools to search the unfamiliar design space for optimum designs. Current methods for incorporating high-fidelity tools into early design phases for the purpose of reducing uncertainty are inadequate due to the severely restricted budgets that are common in early design as well as the unfamiliar design space of advanced aircraft. This motivates the need for a robust and efficient global optimization algorithm. This research presents a novel surrogate model-based global optimization algorithm to efficiently search challenging design spaces for optimum designs. The algorithm searches the design space by constructing a fully Bayesian Gaussian process model through a set of observations and then using the model to make new observations in promising areas where the global minimum is likely to occur. The algorithm is incorporated into a methodology that reduces failed cases, infeasible designs, and provides large reductions in the objective function values of design problems. Results on four sets of algebraic test problems are presented and the methodology is applied to an airfoil section design problem and a conceptual aircraft design problem. The method is shown to solve more nonlinearly constrained algebraic test problems than state-of-the-art algorithms and obtains the largest reduction in the takeoff gross weight of a notional 70-passenger regional jet versus competing design methods.

This thesis explores the possibility of using parallel algorithms to calculate the dynamics of driven quantum systems prevalent in atomic physics. In this process, new as well as existing algorithms are considered. The thesis is split into three parts. In the first part an attempt is made to develop a new formalism of the time dependent Schroedinger equation (TDSE) in the hope that the new formalism could lead to a parallel algorithm. The TDSE is written as an eigenvalue problem, the ground state of which represents the solution to the original TDSE. Even though mathematically sound and correct, it turns out the ground state of this eigenvalue problem cannot be easily found numerically, rendering the original hope a false one. In the second part we borrow a Bayesian global optimisation method from the machine learning community in an effort to find the optimum conditions in different systems quicker than textbook optimisation algorithms. This algorithm is specifically designed to find the optimum of expensive functions, and is used in this thesis to 1. maximise the electron yield of hydrogen, 2. maximise the asymmetry in the photo-electron angular distribution of hydrogen, 3. maximise the higher harmonic generation yield within a certain frequency range, 4. generate short pulses via combining higher harmonics generated by hydrogen. In the last part, the phenomenon of dynamic interference (temporal equivalent of the double-slit experiment) is discussed. The necessary conditions are derived from first principles and it is shown where some of the previous analytical and numerical studies have gone wrong; it turns out the choice of gauge plays a crucial role. Furthermore, a number of different scenarios are presented where interference in the photo-electron spectrum is expected to occur.