Closed-loop optimization deals with problems in which candidate solutions are evaluated by conducting experiments, e.g. physical or biochemical experiments. Although this form of optimization is becoming more popular across the sciences, it may be subject to rather unexplored resourcing issues, as any experiment may require resources in order to be conducted. In this thesis we are concerned with understanding how evolutionary search is affected by three particular resourcing issues -- ephemeral resource constraints (ERCs), changes of variables, and lethal environments -- and the development of search strategies to combat these issues. The thesis makes three broad contributions. First, we motivate and formally define the resourcing issues considered. Here, concrete examples in a range of applications are given. Secondly, we theoretically and empirically investigate the effect of the resourcing issues considered on evolutionary search. This investigation reveals that resourcing issues affect optimization in general, and that clear patterns emerge relating specific properties of the different resourcing issues to performance effects. Thirdly, we develop and analyze various search strategies augmented on an evolutionary algorithm (EA) for coping with resourcing issues. To cope specifically with ERCs, we develop several static constraint-handling strategies, and investigate the application of reinforcement learning techniques to learn when to switch between these static strategies during an optimization process. We also develop several online resource-purchasing strategies to cope with ERCs that leave the arrangement of resources to the hands of the optimizer. For problems subject to changes of variables relating to the resources, we find that knowing which variables are changed provides an optimizer with valuable information, which we exploit using a novel dynamic strategy. Finally, for lethal environments, where visiting parts of the search space can cause the permanent loss of resources, we observe that a standard EA's population may be reduced in size rapidly, complicating the search for innovative solutions. To cope with such scenarios, we consider some non-standard EA setups that are able to innovate genetically whilst simultaneously mitigating risks to the evolving population.

The multi-armed bandit (MAB) problem provides a convenient abstraction for many online decision problems arising in modern applications including Internet display advertising, crowdsourcing, online procurement, smart grids, etc. Several variants of the MAB problem have been proposed to extend the basic model to a variety of practical and general settings. The sleeping multi-armed bandit (SMAB) problem is one such variant where the set of available arms varies with time. This study is focused on analyzing the efficacy of the Thompson Sampling algorithm for solving the SMAB problem. Any algorithm for the classical MAB problem is expected to choose one of K available arms (actions) in each of T consecutive rounds. Each choice of an arm generates a stochastic reward from an unknown but fixed distribution. The goal of the algorithm is to maximize the expected sum of rewards over the T rounds (or equivalently minimize the expected total regret), relative to the best fixed action in hindsight. In many real-world settings, however, not all arms may be available in any given round. For example, in Internet display advertising, some advertisers might choose to stay away from the auction due to budget constraints; in crowdsourcing, some workers may not be available at a given time due to timezone difference, etc. Such situations give rise to the sleeping MAB abstraction. In the literature, several upper confidence bound (UCB)-based approaches have been proposed and investigated for the SMAB problem. Our contribution is to investigate the efficacy of a Thomp-son Sampling-based approach. Our key finding is to establish a logarithmic regret bound, which non-trivially generalizes a similar bound known for this approach in the classical MAB setting. Our bound also matches (up to constants) the best-known lower bound for the SMAB problem. Furthermore, we show via detailed simulations, that the Thompson Sampling approach in fact outperforms the known algorithms for the SMAB problem.

<p>Sequential decision-making is a fundamental class of problem that motivates algorithm designs of online machine learning and reinforcement learning. Arguably, the resulting online algorithms have supported modern online service industries for their data-driven real-time automated decision making. The applications span across different industries, including dynamic pricing (Marketing), recommendation (Advertising), and dosage finding (Clinical Trial). In this dissertation, we contribute fundamental statistical design advances for sequential decision-making algorithms, leaping progress in theory and application of online learning and sequential decision making under uncertainty including online sparse learning, finite-armed bandits, and high-dimensional online decision making. Our work locates at the intersection of decision-making algorithm designs, online statistical machine learning, and operations research, contributing new algorithms, theory, and insights to diverse fields including optimization, statistics, and machine learning.</p> <p><br></p> <p>In part I, we contribute a theoretical framework of continuous risk monitoring for regularized online statistical learning. Such theoretical framework is desirable for modern online service industries on monitoring deployed model's performance of online machine learning task. In the first project (Chapter 1), we develop continuous risk monitoring for the online Lasso procedure and provide an always-valid algorithm for high-dimensional dynamic pricing problems. In the second project (Chapter 2), we develop continuous risk monitoring for online matrix regression and provide new algorithms for rank-constrained online matrix completion problems. Such theoretical advances are due to our elegant interplay between non-asymptotic martingale concentration theory and regularized online statistical machine learning.</p> <p><br></p> <p>In part II, we contribute a bootstrap-based methodology for finite-armed bandit problems, termed Residual Bootstrap exploration. Such a method opens a possibility to design model-agnostic bandit algorithms without problem-adaptive optimism-engineering and instance-specific prior-tuning. In the first project (Chapter 3), we develop residual bootstrap exploration for multi-armed bandit algorithms and shows its easy generalizability to bandit problems with complex or ambiguous reward structure. In the second project (Chapter 4), we develop a theoretical framework for residual bootstrap exploration in linear bandit with fixed action set. Such methodology advances are due to our development of non-asymptotic theory for the bootstrap procedure.</p> <p><br></p> <p>In part III, we contribute application-driven insights on the exploration-exploitation dilemma for high-dimensional online decision-making problems. Such insights help practitioners to implement effective high-dimensional statistics methods to solve online decisionmaking problems. In the first project (Chapter 5), we develop a bandit sampling scheme for online batch high-dimensional decision making, a practical scenario in interactive marketing, and sequential clinical trials. In the second project (Chapter 6), we develop a bandit sampling scheme for federated online high-dimensional decision-making to maintain data decentralization and perform collaborated decisions. These new insights are due to our new bandit sampling design to address application-driven exploration-exploitation trade-offs effectively. </p>