One of the hardest challenges in building a realistic Bayesian network (BN) model is to construct the node probability tables (NPTs). Even with a fixed predefined model structure and very large amounts of relevant data, machine learning methods do not consistently achieve great accuracy compared to the ground truth when learning the NPT entries (parameters). Hence, it is widely believed that incorporating expert judgment or related domain knowledge can improve the parameter learning accuracy. This is especially true in the sparse data situation. Expert judgments come in many forms. In this thesis we focus on expert judgment that specifies inequality or equality relationships among variables. Related domain knowledge is data that comes from a different but related problem. By exploiting expert judgment and related knowledge, this thesis makes novel contributions to improve the BN parameter learning performance, including: • The multinomial parameter learning model with interior constraints (MPL-C) and exterior constraints (MPL-EC). This model itself is an auxiliary BN, which encodes the multinomial parameter learning process and constraints elicited from the expert judgments. • The BN parameter transfer learning (BNPTL) algorithm. Given some potentially related (source) BNs, this algorithm automatically explores the most relevant source BN and BN fragments, and fuses the selected source and target parameters in a robust way. • A generic BN parameter learning framework. This framework uses both expert judgments and transferred knowledge to improve the learning accuracy. This framework transfers the mined data statistics from the source network as the parameter priors of the target network. Experiments based on the BNs from a well-known repository as well as two realworld case studies using different data sample sizes demonstrate that the proposed new approaches can achieve much greater learning accuracy compared to other state-of-theart methods with relatively sparse data.

m-Estimation represents a broad class of estimators, including least-squares and maximum likelihood, and is a widely used tool for statistical inference. Its successful application however, often requires negotiating physical resources for desired levels of accuracy. These limiting factors, which we abstractly refer as costs, may be computational, such as time-limited cluster access for parameter learning, or they may be financial, such as purchasing human-labeled training data under a fixed budget. This thesis explores these accuracy- cost tradeoffs by proposing a family of estimators that maximizes a stochastic variation of the traditional m-estimator. Such "stochastic m-estimators" (SMEs) are constructed by stitching together different m-estimators, at random. Each such instantiation resolves the accuracy-cost tradeoff differently, and taken together they span a continuous spectrum of accuracy-cost tradeoff resolutions. We prove the consistency of the estimators and provide formulas for their asymptotic variance and statistical robustness. We also assess their cost for two concerns typical to machine learning: computational complexity and labeling expense. For the sake of concreteness, we discuss experimental results in the context of a variety of discriminative and generative Markov random fields, including Boltzmann machines, conditional random fields, model mixtures, etc. The theoretical and experimental studies demonstrate the effectiveness of the estimators when computational resources are insufficient or when obtaining additional labeled samples is necessary. We also demonstrate that in some cases the stochastic m-estimator is associated with robustness thereby increasing its statistical accuracy and representing a win-win.

Decision-theoretic systems, such as Markov Decision Processes (MDPs), are used for sequential decision-making under uncertainty. MDPs provide a generic framework that can be applied in various domains to compute optimal policies. This thesis presents techniques that offer explanations of optimal policies for MDPs and then refine decision theoretic models (Bayesian networks and MDPs) based on feedback from experts. Explaining policies for sequential decision-making problems is difficult due to the presence of stochastic effects, multiple possibly competing objectives and long-range effects of actions. However, explanations are needed to assist experts in validating that the policy is correct and to help users in developing trust in the choices recommended by the policy. A set of domain-independent templates to justify a policy recommendation is presented along with a process to identify the minimum possible number of templates that need to be populated to completely justify the policy. The rejection of an explanation by a domain expert indicates a deficiency in the model which led to the generation of the rejected policy. Techniques to refine the model parameters such that the optimal policy calculated using the refined parameters would conform with the expert feedback are presented in this thesis. The expert feedback is translated into constraints on the model parameters that are used during refinement. These constraints are non-convex for both Bayesian networks and MDPs. For Bayesian networks, the refinement approach is based on Gibbs sampling and stochastic hill climbing, and it learns a model that obeys expert constraints. For MDPs, the parameter space is partitioned such that alternating linear optimization can be applied to learn model parameters that lead to a policy in accordance with expert feedback. In practice, the state space of MDPs can often be very large, which can be an issue for real-world problems. Factored MDPs are often used to deal with this issue. In Factored MDPs, state variables represent the state space and dynamic Bayesian networks model the transition functions. This helps to avoid the exponential growth in the state space associated with large and complex problems. The approaches for explanation and refinement presented in this thesis are also extended for the factored case to demonstrate their use in real-world applications. The domains of course advising to undergraduate students, assisted hand-washing for people with dementia and diagnostics for manufacturing are used to present empirical evaluations.

Decision-theoretic systems, such as Markov Decision Processes (MDPs), are used for sequential decision-making under uncertainty. MDPs provide a generic framework that can be applied in various domains to compute optimal policies. This thesis presents techniques that offer explanations of optimal policies for MDPs and then refine decision theoretic models (Bayesian networks and MDPs) based on feedback from experts. Explaining policies for sequential decision-making problems is difficult due to the presence of stochastic effects, multiple possibly competing objectives and long-range effects of actions. However, explanations are needed to assist experts in validating that the policy is correct and to help users in developing trust in the choices recommended by the policy. A set of domain-independent templates to justify a policy recommendation is presented along with a process to identify the minimum possible number of templates that need to be populated to completely justify the policy. The rejection of an explanation by a domain expert indicates a deficiency in the model which led to the generation of the rejected policy. Techniques to refine the model parameters such that the optimal policy calculated using the refined parameters would conform with the expert feedback are presented in this thesis. The expert feedback is translated into constraints on the model parameters that are used during refinement. These constraints are non-convex for both Bayesian networks and MDPs. For Bayesian networks, the refinement approach is based on Gibbs sampling and stochastic hill climbing, and it learns a model that obeys expert constraints. For MDPs, the parameter space is partitioned such that alternating linear optimization can be applied to learn model parameters that lead to a policy in accordance with expert feedback. In practice, the state space of MDPs can often be very large, which can be an issue for real-world problems. Factored MDPs are often used to deal with this issue. In Factored MDPs, state variables represent the state space and dynamic Bayesian networks model the transition functions. This helps to avoid the exponential growth in the state space associated with large and complex problems. The approaches for explanation and refinement presented in this thesis are also extended for the factored case to demonstrate their use in real-world applications. The domains of course advising to undergraduate students, assisted hand-washing for people with dementia and diagnostics for manufacturing are used to present empirical evaluations.

Variational methods and Partial Differential Equations (PDEs) have been extensively employed for the mathematical formulation of a myriad of problems describing physical phenomena such as heat propagation, thermodynamic transformations and many more. In imaging, PDEs following variational principles are often considered. In their general form these models combine a regularisation and a data fitting term, balancing one against the other appropriately. Total variation (TV) regularisation is often used due to its edgepreserving and smoothing properties. In this thesis, we focus on the design of TV-based models for several different applications. We start considering PDE models encoding higher-order derivatives to overcome wellknown TV reconstruction drawbacks. Due to their high differential order and nonlinear nature, the computation of the numerical solution of these equations is often challenging. In this thesis, we propose directional splitting techniques and use Newton-type methods that despite these numerical hurdles render reliable and efficient computational schemes. Next, we discuss the problem of choosing the appropriate data fitting term in the case when multiple noise statistics in the data are present due, for instance, to different acquisition and transmission problems. We propose a novel variational model which encodes appropriately and consistently the different noise distributions in this case. Balancing the effect of the regularisation against the data fitting is also crucial. For this sake, we consider a learning approach which estimates the optimal ratio between the two by using training sets of examples via bilevel optimisation. Numerically, we use a combination of SemiSmooth (SSN) and quasi-Newton methods to solve the problem efficiently. Finally, we consider TV-based models in the framework of graphs for image segmentation problems. Here, spectral properties combined with matrix completion techniques are needed to overcome the computational limitations due to the large amount of image data. Further, a semi-supervised technique for the measurement of the segmented region by means of the Hough transform is proposed.