Author: Inga Catharina Ibs

Referees:
Prof. Constantin Rothkopf, Ph.D.
Prof. Dr. Ute Schmid

Defense: 27.05.2026

Abstract:

Humans engage in a wide variety of tasks, ranging from preparing breakfast to playing games such as Tetris to planning resource allocations in energy generation, all the way to sketching out long-term life paths.

What all these tasks have in common is that they involve sequential decision-making, \ie, early decisions significantly affect the options available later, and the final outcome remains unclear throughout much of the decision-making process. The diversity of tasks not only results from differences in their objectives but also from other features that describe each problem's structure, such as the number of variables or whether the relationships among variables, actions, and outcomes are static or dynamic. This diversity in sequential decision-making tasks is reflected in the landscape of available computational solutions. While some resource allocation problems can be solved exactly, \eg, using linearly constrained optimization, games such as chess and Go have only approximate decision algorithms.

Understanding human behavior in sequential decision-making tasks requires investigating both the computational properties of the underlying problem and humans' cognitive representations and solution approaches. Structural features vary considerably across tasks and play a central role because they directly shape how humans approach problem solving, and simultaneously constrain which algorithms can achieve optimal solutions, thereby providing a normative baseline for cognitive modeling. While the tasks used classically to study human sequential decision-making are designed to be paradigmatic of real-world problems, they often isolate specific task features to ease the computational descriptions of cognitive processes. Therefore, models that capture how people solve a specific task are not easily transferred to other tasks or to real-world problems involving the full breadth of computational task complexities.

Even if optimal solutions for a specific sequential decision-making problem are available, understanding human behavior may still be challenging. The ability of computers to find solutions to complex sequential decision-making problems in a reasonable amount of time often depends on mathematical abstractions of the original problem, which bear little resemblance to how people naturally represent such problems. As a result, computational solutions often remain opaque, even to experts, a fundamental problem relevant to Explainable AI. Humans may employ very different internal representations of a decision-making problem. Thus, they may use very different algorithms to solve a specific problem, which also influences how they reason about solutions. One way to bridge the gap between computational solutions and human behavior is to understand how humans explain them. This approach also offers the opportunity to ground explanations for algorithmic solutions in an understanding of humans' cognitive representations, sequential decisions, and reasoning about the problem at hand.

In this thesis, we adopt a task-centric perspective on sequential decision-making to gain a deeper understanding of human behavior in such tasks. Specifically, we investigate how sequential decision-making tasks can be systematized based on their structural features and how these structural task features impact human problem-solving behavior. We further investigate how human problem-solving behavior can be understood in terms of underlying mental representations and how these representations can provide a foundation for generating explanations of computational solutions.

We begin by proposing a taxonomy of sequential decision-making tasks based on structural features that determine the computational goals optimal algorithmic solutions must account for. This taxonomy provides a systematic framework for understanding the representations and strategies people might employ across different tasks. Based on this taxonomy, we compare experimental tasks and identify groups of tasks with similar feature combinations.

We continue by studying demonstration-based reward learning as a domain where cognitive models, such as Bayesian inverse reinforcement learning, have successfully captured human behavior. We extend previous investigations by examining how people's strategies adapt to differences in task demands, specifically in terms of task features such as environmental complexity and information availability. Using explicit behavioral measures, \ie, sequential decisions, and implicit measures, in the form of eye movements, we demonstrate that individuals shift between model-based inference and more approximate or biased inference strategies. We propose extensions to the modeling framework to account for these context-dependent adaptations.

Next, we examine deterministic optimization tasks involving multidimensional transitions as a representative case of complex sequential decision-making without uncertainty. To investigate human mental representations and problem solving in this class of problems, we developed a new experimental task of a resource allocation problem in the form of a game, the Furniture Factory. We use this experimental task to investigate human problem-solving strategies across three studies, using both qualitative and quantitative methods. First, we elicit concurrent explanations from participants in a think-aloud setting as they describe their decisions. We analyse these explanations and their decisions, and identify the representations and heuristics underlying them. In a second study, we use a sequential decision-making version of the Furniture Factory and elicit post-hoc explanations, which we relate to the representations found in the first study. Based on the results of the two studies, we formalize strategies that, when combined, can serve as descriptors of participants' behavior and optimal solutions. We validate these formal strategies on a large behavioral dataset from a third experiment.

The optimization problems for which we elicited explanations share the same task features as problems in the domain of linear constrained optimization, which occur in many real-world applications. Although solution methods for such problems are theoretically well understood, it remains unclear what form intuitive, human-understandable explanations of these solutions should take. Building on insights from our experiments, we propose an algorithm for finding rationales that align with human representations of constrained optimization problems. We base the algorithm on a grammar of predicates derived directly from participants' explanations from our previous studies on human optimization strategies. This grammar forms the foundation of a rational rules model, which we use to generate rationales modeled after human explanations. We evaluate our algorithm on human solutions and demonstrate that the generated rationales capture human decision processes well and align with the representations and structure of human explanations.

In summary, this thesis contributes to a deeper understanding of sequential decision-making by exploring the relationship between tasks based on their structural features. We take these structural features into account when we explore human decision-making in the domains of constrained optimization and learning from action demonstrations using a variety of experimental paradigms and cognitive process-tracing methods. Finally, we contribute to the field of explainable AI by showing how insights into human problem-solving strategies can be used to generate human-aligned explanations of computational solutions.