Abstract
Understanding the hierarchical nature of human planning has traditionally been limited by the inability to observe internal mental structures directly. This study utilizes the programming game Lightbot as a process-tracing tool to make these latent hierarchies explicit. By analyzing the programs written by 193 participants across ten navigation tasks, the researchers compared models of planning based on step-cost minimization, program compression (Minimum Description Length), and grammar induction. The findings reveal a robust human bias toward subroutine reuse, suggesting that human planning is better modeled as a language-like process of building a vocabulary of behavior rather than a simple optimization of action sequences.
Introduction
In everyday life, human planning is inherently hierarchical. Complex tasks like preparing a meal are decomposed into sub-goals, which are further broken down into specific motor actions. However, cognitive scientists have historically struggled to study these hierarchies because they are typically unobservable; researchers usually only see the final, flat sequence of actions. Traditional methods have relied on indirect proxies like temporal pauses or error rates to infer structure. This research introduces a "process-tracing" approach by asking participants to write actual programs to solve tasks, thereby making their mental blueprints for action explicit.
Methodology
The study employed Lightbot, an educational game where players must program a robot to navigate a gridworld and activate specific squares. The interface allows for the creation of a "Main" program and several subroutines (Processes), providing a direct window into how individuals "chunk" actions into reusable units.
Modeling Framework
To explain how humans choose specific program structures, the authors compared three computational theories:
- Step Cost: A utility-maximization model where humans aim to minimize the total number of physical actions the robot performs.
- Minimum Description Length (MDL): A compression-based model where humans minimize the length of the program code itself.
- Grammar Induction: A novel approach based on Adaptor Grammars. This model treats planning like language acquisition, using a Dirichlet Process to create a "rich-get-richer" effect. Once a subroutine is established, it becomes more likely to be reused, even if it does not result in the shortest possible program.
Data Collection and Analysis
The researchers recruited 193 participants via Prolific to solve 10 Lightbot tasks of varying complexity, resulting in a dataset of 1,668 programs. To facilitate model comparison, a preprocessing pipeline was used to "canonicalize" programs (e.g., inlining subroutines used only once), which significantly improved the models' ability to account for human solutions.
Key Results
The Dominance of Grammar Induction
The Grammar Induction + Step Cost model emerged as the superior fit, achieving the lowest Bayesian Information Criterion (BIC) across nearly all tasks. This suggests that human planning is not merely about minimizing effort or code length, but about inducing a structured "grammar" of actions.
The Reuse Bias and Incentive Paradox
A striking finding was the "Reuse Bias": participants frequently chose programs with high subroutine reuse even when those programs were longer or less efficient than available alternatives. Remarkably, this bias persisted even when participants were offered financial bonuses to produce the shortest possible programs (MDL). This suggests that the preference for elegant, reusable structures is a fundamental feature of human cognition.
Cognitive Benefits of Hierarchy
The study found a significant negative correlation between program "commonality" (the degree to which a program followed common reuse patterns) and task difficulty. Participants who utilized common, high-reuse programs exhibited faster response times ($\beta = -0.008, p < .001$) and required fewer trial-and-error evaluations ($\beta = -0.053, p < .001$). This indicates that a well-structured hierarchy makes the plan easier for the human agent to reason about and execute.
Discussion and Future Work
While the Lightbot interface may provide a scaffold that encourages hierarchical thinking, the results strongly support the view of planning as a language-like induction process. Future research should investigate whether humans maintain "libraries" of subroutines across different tasks and explore the specific algorithmic processes the brain uses to search the vast space of possible programs. Expanding this research into more naturalistic, non-gridworld settings will be essential to confirming the universality of the reuse bias in human planning.
Summary updated