Intuitive experimentation in the physical world

Bramley et al., 2018

Source: Bramley et al., 2018


  • How do people actively experiment to discover latent properties within the physical world?
  • Performed experiments where participants interacted with objects in order to identify the masses or attractive/repulsive forces that governed their movement
  • The interactions selectively produced evidence that revealed the physical property being investigated
    • Active learners produced more accurate inferences than passive and yoked observers
  • Links: [ website ]


  • From a causal perspective, actions can be seen as interventions which help reveal how to world works
    • Unclear how these actions are chosen, and how complex and informative they are
  • Active learning studies in humans usually focus on differentiating between a relatively small number of discrete hypotheses
    • Evidence for its utility has also been mixed
  • This study explores active learning in a continuously dynamic environment
    • Learner must decide where, how, and when to intervene
  • Prior work using a similar simulated environment demonstrated that humans are worse at detecting attraction versus repulsion when passively observing
    • Divergence corresponds to asymmetry in evidence, since objects that repel each other rarely spend enough time close together


  • Aim to establish how effective participants’ actions are at reducing uncertainty about the properties they’re asked about
  • 2D microworld environment - contains a number of colored pucks with varying masses and local (magnet-like) pair-wise forces
    • Active: allow participants to grab objects by clicking on them and dragging
    • Passive: participants only observe the world
    • Yoked: passively observe actions of active participant
  • Ideal Observer (IO): predicts how the scene would unfold given various assumptions about the properties and comparing to whats observed
    • Predictive Divergence (PD): take divergence between simulations varying the property and what is observed, averaging over all possible settings for other properties


  • Just looking at accuracy, active participants were significantly better than passive and yoked participants, which performed similarly
    • Active participants’ performance was moderately predictive of their yoked counterparts’ (r=0.49)
  • In general, participants were worse at inferring masses compared to forces
    • Active learners benefited the most for inferring repulsion
    • Yoked participants had marginal improvement in inferring forces
  • While active participants generated the same amount of evidence overall as passive participants, they generated more evidence about force
  • PD for force and mass of target was higher for active participants compared to passive, also for intervening periods of active participants compared to observing periods
  • Giving the same learning goal to yoked participants raises their performance to the level of their active counterparts


  • Provides additional evidence for the importance of active learning and volitional control of interventions
  • How these sophisticated strategies are learned remains an open question
Elias Z. Wang
Elias Z. Wang
AI Researcher | PhD Candidate