Explore, Exploit, Create: Quantifying the dynamics of exploratory play

• In Event: Poster Session 14
  In Poster Session: PS 14 Topic - Cognitive Processes

Abstract

Play is a universal human behavior widely held to be critical for learning and development. To understand how play contributes to learning, recent work has compared children’s exploratory play with formal accounts of rational learning. This “play as rational exploration” view suggests that children’s play is sensitive to costs, rewards, and expected information gain. However, one of the most salient features of children’s play is its richness and variability. Further, the dynamics of free play remains unspecified even though there exist precise and well-validated models of rational exploration, which predict that learners switch between exploring novel options for unknown rewards and exploiting familiar options for known rewards. The dynamic trade-off between these two behavioral phases has been successfully captured and modeled in a variety of human, animal, and computational studies (for review, see Wilson et al, 2020).

This project aims to ask if children’s exploratory play follows this signature dynamic and to obtain a more complete description of play by considering additional drivers of play behavior. Specifically we propose that in play children not only trade-off between exploring new rewards and exploiting known rewards but additionally make up their own rewards by defining their own goals -- goals which may be arbitrary and have no direct relation to information gain or other environmentally-defined rewards. We designed a novel free play task, asking participants to play on an 8x8 grid of buttons for 5 minutes (see Figure 1) with no explicit instruction on what to do. When activated, the buttons turn blue and play one of four sound effects that vary in reward (e.g. a short “blip” versus a pleasant “ring”). Participants can end the play session at their leisure. We recorded participants’ clicks and generated real-time videos of participant behavior. Trained coders will annotate these videos and label behavioral segments as either exploration, exploitation, or creation, with several sub-categories.

Our hypothesis is that play can be characterized as three phases each with their own dynamic trajectories. As time progresses and less novel information can be gained, the reward value of exploration will diminish and children will be more likely to exploit interesting affordances. As these are repeated, the reward of exploitation will begin declining. At this point we expect children to shift to more costly creation goals where participants seek to achieve certain end states (e.g. nameable objects, common symbols, symmetric patterns, etc.) and design plans accordingly. We propose to develop a computational model that can capture the dynamics of children’s play, including these behavioral phases and how they explore the 2^64 possible goal states. This computational model allows us to make quantitative evaluations of developmental changes in the nature of play and creativity, including variability, efficiency, and goal-directedness.

Currently we have data from adult participants (N=120) and are actively recruiting child participants (target N = 120, ages 5-8 years). We expect to finish data collection and behavior coding by December 2020, leaving sufficient time for data analysis and evaluating computational models.

Authors

• Sophia Diggs-Galligan, Massachusetts Institute of Technology (MIT)
  Presenting Author

• Junyi Chu, Massachusetts Institute of Technology (MIT)
  Non-Presenting Author

• Joshua B. Tenenbaum, Massachusetts Institute of Technology
  Non-Presenting Author

• Laura E Schulz, Massachusetts Institute of Technology (MIT)
  Non-Presenting Author