Bad decisions can have devastating consequences, and there is a vast body of literature suggesting that human judgment and decision-making are riddled with numerous systematic violations of the rules of logic, probability theory, and expected utility theory. The discovery of these cognitive biases in the 1970s challenged the concept of Homo sapiens as the rational animal and has profoundly shaken the foundations of economics and rational models in the cognitive, neural, and social sciences. Four decades later, these disciplines still lack a rigorous theoretical foundation that can account for people’s cognitive biases. Furthermore, designing effective interventions to remedy cognitive biases and improve human judgment and decision-making is still an art rather than a science. I address these two fundamental problems in the first and the second part of my thesis respectively.

To develop a theoretical framework that can account for cognitive biases, I start from the assumption that human cognition is fundamentally constrained by limited time and the human brain’s finite computational resources. Based on this assumption, I redefine human rationality as reasoning and deciding according to cognitive strategies that make the best possible use of the mind’s limited resources. I draw on the bounded optimality framework developed in the artificial intelligence literature to translate this definition into a mathematically precise theory of bounded rationality called resource-rationality and a new paradigm for cognitive modeling called resource-rational analysis. Applying this methodology allowed me to derive resource-rational models of judgment and decisionmaking that accurately capture a wide range of cognitive biases, including the anchoring bias and the numerous availability biases in memory recall, judgment, and decision-making. By showing that these phenomena and the heuristics that generate them are consistent with the rational use of limited resources, my analysis provides a rational reinterpretation of cognitive biases that were once interpreted as hallmarks of human irrationality. This suggests that it is time to revisit the debate about human rationality with the more realistic normative standard of resource-rationality. To enable a systematic assessment of the extent to which human cognition is resource- rational, I present an automatic method for deriving resource-rational heuristics from a mathematical specification of their function and the mind’s computational constraints. Applying this method to multi-alternative risky-choice led to the discovery of a previously unknown heuristic that people appear to use very frequently. Evaluating human decision-making against resource-rational heuristics suggested that, on average, human decision-making is at most 88% as resource-rational as it could be.

Since people are equipped with multiple heuristics, a complete normative theory of bounded rationality also has to answer the question of when each of these heuristics should be used. I address this question with a rational theory of strategy selection. According to this theory, people gradually learn to select the heuristic with the best possible speed-accuracy trade-off by building a predictive model of its performance. Experiments testing this model confirmed that people gradually learn to make increasingly more rational use of their finite time and bounded cognitive resources through a metacognitive reinforcement learning mechanism.

Overall, these findings suggest that–contrary to the bleak picture painted by previous research on heuristics and biases–human cognition is not fundamentally irrational, and can be understood as making rational use of bounded cognitive resources. By reconciling rationality with cognitive biases and bounded resources, this line of research addresses fundamental problems of previous rational modeling frameworks, such as expected utility theory, logic, and probability theory. Resource-rationality might thus come to replace classical notions of rationality as a theoretical foundation for modeling human judgment and decision-making in economics, psychology, neuroscience, and other cognitive and social sciences.

In the second part of my dissertation, I apply the principle of resource-rationality to develop tools and interventions for improving the human mind. Early interventions educated people about cognitive biases and taught them the normative principles of logic, probability theory, and expected utility theory. The practical benefits of such interventions are limited because the computational demands of applying them to the complex problems people face in everyday life far exceed individuals’ cognitive capacities. Instead, the principle of resource-rationality suggests that people should rely on simple, computationally efficient heuristics that are well adapted to the structure of their environments. Building on this idea, I leverage the automatic strategy discovery method and insights into metacognitive learning from the first part of my dissertation to develop intelligent systems that teach people resource-rational cognitive strategies. I illustrate this approach by developing and evaluating a cognitive tutor that trains people to plan resource-rationally. My results show that practicing

with the cognitive tutor improves people’s planning strategies significantly more than does practicing without feedback. Follow-up experiments demonstrate that this training effect transfers to more difficult planning problems in novel and more complex environments, and that this transfer effect is retained over time. This indicates that discovering and teaching resource-rational heuristics may

be a promising approach to improving human judgment and decision-making. While this approach adapts people’s heuristics to the structure of their environment, the theory of resource-rationality suggests that human decision-making can also be improved by adapting the structure of the environment to the heuristics people already use. I illustrate this approach by developing a cognitive prosthesis for goal achievement that helps people overcome procrastination, spring into action, and achieve their goals on time.

By virtue of integrating rational principles with cognitive constraints, resource-rationality provides a realistic normative standard for human reasoning and decision-making. My findings about human rationality and metacognitive learning are consistent with the view that evolution and learning adapt the mind to the structure of its environment and the constraints imposed by its limited resources. These adaptive mechanisms appear to optimize for resource-rationality, and the benefits of training with the cognitive tutor demonstrate that this adaptation can be accelerated with the help of artificial intelligence. This makes resource-rationality a promising theoretical framework for modeling and improving human cognition.

The greatest human achievements are never completed alone. However, social interaction is complex and successful collaboration even more so. Tremendous amounts of information and processing are involved in predicting, interpreting, and working with others. These calculations are implicit, deeply embedded in the human psyche and not easily accessible for analysis or improvement. Knowing what algorithms the mind is solving when collaborating with others would be invaluable, both for our own knowledge and improvement of social collaboration between people, and to reconstruct these algorithms in systems outside ourselves. People increasingly interact with artificial intelligence systems, and developing models of social interaction could enable these systems to seamlessly assist us. This premise motivates my work: that by understanding the algorithms behind social collaboration, we can improve interactions between people, and between people and machines.

In this dissertation, I formalize and test computational models of collaboration, focusing on three problem areas. First, I investigate how people collaborate in recalling information. Though one might expect memory to be a solitary venture, researchers have long studied the differences in how people recall information in groups compared to alone, though only for small group sizes. Luhmann and Rajaram (2015) hypothesized mechanisms of large-scale recall using an agent-based model. I test the predictions of this model with an empirical experiment recruiting thousands of participants. Second, I investigate another key component of collaboration: people’s intuitive judgments of how shared resources should be allocated among people with different preferences. I collect empirical data from participants under a number of decision conditions, then use an inverse reinforcement learning model to determine what underlying mathematical fairness principles characterize people’s choices. Third, I investigate how people infer others’ preferences, a key component of collaboration. My coauthors and I present a rational model for inferring preferences from response times, using a Drift Diffusion Model to characterize how preferences influence response time and Bayesian inference to invert this relationship. We then compare the model’s predictions to collected empirical data. These three case studies comprise novel models of social interaction, tested with behavioral experiments, aimed at improving human and technological collaboration.