# Exploring the hierarchical structure of human plans via program generation

**Authors:** Carlos G. Correa, Sophia Sanborn, Mark K. Ho, Frederick Callaway, Nathaniel D. Daw, Thomas L. Griffiths  
**arXiv:** [2311.18644](https://arxiv.org/abs/2311.18644)  
**Date:** 2023-11-30  
**Summary page:** https://www.summarizepaper.com/en/paper/2311.18644v1/

## Key points

- The paper explores the hierarchical structure of human plans
- Participants were asked to create programs with explicit hierarchical structure
- Two principles of human behavior were tested: utility maximization and minimum description length
- Humans are sensitive to both metrics, but there is a preference for programs with reuse beyond what MDL predicts
- The researchers extended the MDL account into a generative model over programs, modeling hierarchy choice as the induction of a grammar over actions
- This extended account provides the best prediction of human behavior and highlights a principle that guides hierarchical planning
- The paper draws on literatures such as sequence learning, reinforcement learning, and Bayesian program induction to inform their approach

## Summary

This paper explores the hierarchical structure of human plans by using an experimental paradigm that allows for the observation of hierarchical representations. The participants in the study were asked to create programs that generate sequences of actions in a language with explicit hierarchical structure. The researchers aimed to test two well-established principles of human behavior: utility maximization (using fewer actions) and minimum description length (having a shorter program). They found that humans are sensitive to both metrics, but neither account could predict a qualitative feature of human-created programs, which is the preference for programs with reuse beyond what MDL predicts. To explain this preference, the researchers extended the MDL account into a generative model over programs, modeling hierarchy choice as the induction of a grammar over actions. This extended account provides the best prediction of human behavior and highlights a principle that guides hierarchical planning. The paper draws on literatures such as sequence learning, reinforcement learning, and Bayesian program induction to inform their approach to modeling complex, hierarchical planning.