Robust and Reliable Decision-Making Under Uncertainty

Sequential decision-making is a fundamental problem encountered in many application areas such as robotics, finance, and healthcare. Since many of these decision-making problems are based on data, they are inherently affected by uncertainty that arises from insufficient, incomplete, or incorrect data. With the rise of artificial intelligence (AI), the reliance on data to solve decision-making problems is greater than ever, as also evidenced by the success of reinforcement learning (RL). As a consequence, uncertainty in sequential decision-making is unavoidable. This uncertainty must be accounted for to ensure solutions to these decision-making problems are robust and reliable against incomplete information, misspecification, or adversarial perturbation while maintaining performance. This PhD thesis presents several novel algorithms and techniques to increase the robustness and reliability of AI systems used to solve sequential decision-making problems.