Action subsampling supports policy compression in large action spaces.

Real-world decision-making necessitates navigating expansive action spaces where state-dependent values impose significant cognitive demands on the individual. Prior research has shown that humans face inherent limitations when generating action consideration sets or refining complex state-action mappings. These internal constraints often force agents to simplify multifaceted problems to fit within restricted processing capacities available for neural computation. While scientists have examined these two distinct bottlenecks independently, the specific relationship between reduced consideration sets and policy complexity remains poorly understood. Understanding how these factors interact is essential for modeling realistic human behavior in environments characterized by high dimensionality and uncertainty. The lack of a cohesive model prevents a full understanding of how cognitive load alters the selection of viable options. This absence of evidence motivated the development of a unified perspective on decision-making under cognitive limitations.

This research integrates action consideration sets and policy complexity into a single resource-rational framework for policy compression. The study seeks to provide a comprehensive explanation for how humans manage large action spaces under severe cognitive load by balancing different types of internal costs. Researchers aimed to characterize the suboptimality that arises when individuals reduce the number of options they actively consider during the deliberation process. By simulating various decision-making scenarios, the team investigated the complex interaction between policy simplification and the size of the consideration set. The work also attempts to explain empirical phenomena such as preferential sampling of high-value options across different environmental contexts. The project also evaluates whether humans employ adaptive metacognitive strategies to maintain performance in task-dependent environments where resources are scarce. Finally, the authors sought to validate their computational predictions through direct observation of human behavior in controlled experimental settings.

The investigators developed a computational model based on resource-rational principles to simulate decision-making processes within high-dimensional action spaces. They utilized these simulations to measure the suboptimality resulting from restricted action consideration sets across varying levels of policy complexity. To validate the theoretical predictions, the team conducted a contextual Multi-Armed Bandit (MAB) experiment involving human participants who performed tasks under varying conditions. This experimental setup required subjects to select actions from a large set while their cognitive resources were monitored through performance metrics. The researchers analyzed response correlations across different contexts to identify patterns of preferential sampling and option generation. Statistical frameworks were applied to compare human performance against the normative benchmarks established by the initial simulation results. This approach allowed the team to determine how participants flexibly adjusted their internal parameters to meet task demands.

Simulations revealed that reducing the size of action consideration sets significantly increases suboptimality unless the agent compensates with specific policy adjustments. The interaction between policy complexity and the number of considered options proved to be a functional determinant in mitigating performance loss during complex tasks. Empirical data from the multi-armed bandit (MAB) task showed that humans flexibly adapt their consideration sets based on environmental demands and available resources. Participants demonstrated increased correlation in their responses across different contexts when placed under high cognitive load, reflecting a simplification of their internal models. The findings confirmed that individuals preferentially sample options with high general value when their processing capacity is severely limited. These results suggest that human decision-makers maintain near-optimality through task-dependent adjustments of their internal policy structures and action sets. The data supported the hypothesis that resource-rational trade-offs govern the compression of policies in large-scale environments.

The study demonstrates that accounting for fine-grained resource constraints is vital for understanding the architecture of human cognition in real-world scenarios. These findings highlight the existence of sophisticated metacognitive strategies that allow for efficient decision-making in both simple and complex tasks. The proposed framework provides a robust tool for predicting how cognitive load influences option generation and state-action mappings across diverse domains. Future research may apply these principles to broader models of human behavior where decision-making processes are constrained by limited processing power. By framing policy compression as a resource-rational trade-off, the authors offer a new lens for viewing human adaptability in the face of complexity. This work underscores the importance of integrating multiple cognitive bottlenecks into unified computational models to capture the nuances of human behavior. Ultimately, the research suggests that human intelligence is defined by the ability to compress information without sacrificing significant utility.