Intuitive experimentation in the physical world

Neil R Bramley, Tobias Gerstenberg, Joshua B. Tenenbaum, Todd M. Gureckis

Abstract

Humans possess a remarkable ability to uncover hidden physical properties of objects through active interaction. This paper investigates how individuals act as "intuitive scientists" by performing targeted micro-experiments to resolve uncertainty about latent variables like mass and magnetic force. Using a physics-engine-based "microworld" and a computational Ideal Observer model, the research demonstrates that human interventions are strategically designed to maximize information gain and filter environmental noise.

Introduction

While visual cues provide immediate information about an object's color or shape, fundamental physical properties such as mass or magnetic attraction are often "latent"—hidden from direct observation. To understand these properties, humans do not merely watch; they interact. We lift, shake, and collide objects to probe their underlying nature. This study, published in Cognitive Psychology, explores the cognitive mechanisms behind these "micro-experiments" and how they align with formal principles of active learning.

The Problem of Active Learning in Dynamic Environments

Previous research into active learning has largely focused on discrete, symbolic tasks (such as the Wason selection task). However, the physical world is continuous, dynamic, and often chaotic. Determining how a human decides to intervene in a complex system to resolve uncertainty requires a bridge between high-level reasoning and low-level physical dynamics. The authors address this gap by examining how people choose specific actions to uncover hidden properties in a simulated environment.

Methodology: Analysis by Synthesis

The researchers utilized a digital "microworld" powered by the Box2D physics engine. Participants observed and interacted with circular pucks on a frictionless digital surface. To evaluate human performance, the study employed an Ideal Observer (IO) model based on the "analysis by synthesis" framework.

The IO model maintains multiple hypotheses about the world (e.g., "Puck A is twice as heavy as Puck B") and runs internal simulations for each. By comparing these simulations to observed motion, the model updates its beliefs. A key innovation is the metric of Predictive Divergence (PD). Unlike standard uncertainty reduction, PD is prospective—it measures the degree to which future states of the system are expected to differ based on a specific property. High PD indicates that an action is likely to yield highly discriminative information, allowing the observer to distinguish between competing hypotheses.

Experimental Design and Data Collection

The study involved two primary experiments. Experiment 1 tasked 64 participants with identifying both relative mass and force relationships (attraction/repulsion) between pucks. Experiment 2 (120 participants) used single-goal trials to determine if strategies shifted based on the specific property being investigated.

Participants were divided into three groups:

  1. Active Learners: Could freely manipulate pucks using a mouse.
  2. Passive Observers: Watched the pucks move under natural dynamics.
  3. Yoked Observers: Watched recordings of active learners' sessions.

The researchers compiled a rich behavioral dataset of 2,313 individual human interventions, which were manually coded to categorize specific strategic behaviors.

Key Insights: The Taxonomy of Micro-Experiments

Analysis of the coded interventions revealed that humans employ specific, non-random strategies tailored to the property of interest:

  • Shaking: Rapidly oscillating a puck to gauge its inertia, primarily used to determine mass.
  • Encroaching: Bringing pucks into close proximity to trigger latent forces, essential for identifying magnetism.
  • Deconfounding: Moving irrelevant "distractor" pucks away to isolate the interaction between target objects.

A significant finding was the "repulsion advantage." In passive settings, repulsive forces are difficult to identify because objects naturally move apart, ending the interaction quickly. Active learners overcome this by "forcing" pucks together, creating high-evidence scenarios that would never occur spontaneously.

Key Results

Active learners significantly outperformed passive observers across all conditions. In Experiment 1, active learners identified repulsion with 90% accuracy, compared to only 53% for passive observers. Furthermore, the study identified a "yoking gap": observers watching an active learner were only effective if they shared the same goal. If an active learner was shaking a puck to test mass, but the yoked observer was trying to learn about force, the observer failed to gain meaningful information. This highlights the importance of "first-hand" knowledge of experimental intent in interpreting physical data.

The specificity of human strategies was also statistically significant. When asked about mass, participants predominantly "shook" pucks; when asked about force, they "encroached." The IO model confirmed that these actions generated significantly higher Predictive Divergence for the target property than for non-target properties. Notably, two strategies—"Shaking" and "Throwing"—explained 58% of the variance in participants' accuracy across mass trials.

Discussion and Future Directions

The research suggests that human physical interaction is formally effective at maximizing information gain. However, several questions remain. The current study lacked haptic feedback, which in the real world provides direct sensory data on weight and resistance. Additionally, while the IO model uses full simulation, humans might rely on simpler heuristics or "rules of thumb." Future research could explore the developmental trajectory of these skills—investigating when children transition from simple exploratory play to the sophisticated, goal-directed experimentation seen in adults.

Conclusion

By framing human interaction as a form of scientific inquiry, this research provides a powerful framework for understanding how we navigate and learn about the physical world. These insights have broad implications, from improving STEM education to developing more intuitive and capable robotic systems.

Reference: Bramley, N. R., Gerstenberg, T., Tenenbaum, J. B., & Gureckis, T. M. (2018). Intuitive experimentation in the physical world. Cognitive Psychology, 105, 9-38.

Summary updated

Abstract

Many aspects of our physical environment are hidden. For example, it is hard to estimate how heavy an object is from visual observation alone. In this paper we examine how people actively "experiment" within the physical world to discover such latent properties. In the first part of the paper, we develop a novel framework for the quantitative analysis of the information produced by physical interactions. We then describe two experiments that present participants with moving objects in "microworlds" that operate according to continuous spatiotemporal dynamics similar to everyday physics (i.e., forces of gravity, friction, etc...). Participants were asked to interact with objects in the microworlds in order to identify their masses, or the forces of attraction/repulsion that governed their movement. Using our modeling framework, we find that learners who freely interacted with the physical system selectively produced evidence that revealed the physical property consistent with their inquiry goal. As a result, their inferences were more accurate than for passive observers and, in some contexts, for yoked participants who watched video replays of an active learner's interactions. We characterize active learners' actions into a range of micro-experiment strategies and discuss how these might be learned or generalized from past experience. The technical contribution of this work is the development of a novel analytic framework and methodology for the study of interactively learning about the physical world. Its empirical contribution is the demonstration of sophisticated goal directed human active learning in a naturalistic context.
BibTeX
@inproceedings{bramley2018intuitive,
  author = {Neil R Bramley and Tobias Gerstenberg and Joshua B. Tenenbaum and Todd M. Gureckis},
  title = {{Intuitive experimentation in the physical world}},
  booktitle = {Cognitive},
  year = {2018},
  doi = {10.1016/j.cogpsych.2018.05.001},
}
Figure