Developing Intelligence

Play is more often simply observed than studied scientifically – play behaviors occur unpredictably and, when they do occur, are highly chaotic, making it very difficult to study them in the laboratory. Despite these challenges, new work is beginning to make play accessible from a rigorous scientific framework.

For example, a recent article by Schulz & Bonawitz takes Piaget’s notion of play as a mechanism for understanding causal relationships and recasts it into a testable prediction: children should be more likely to play with an object about which they have incomplete or confounded evidence.

To test the idea, Schulz & Bonawitz carefully crafted two toys in the form of a simplified jack-in-the-box: each toy consisted of a box with a puppet inside; the puppet could be made to emerge from a hole in the top of the box by moving a single lever on the box’s side. The important difference between these toys was the number of levers: one of the boxes had an additional lever on the side which controlled an additional puppet. An experimenter introduced each child to the two-lever toy first, and then gave them the opportunity to continue playing with that toy, or to switch to the single-lever toy.


Those children who hadn’t been able to determine the causal structure of the first familiar toy – i.e., the child and experimenter were always coordinated in depressing the two levers, so that children couldn’t determine which lever controlled which toy – were more likely to continue playing with it when given the opportunity than three other control conditions, where the evidence was disambiguated. (These control conditions were included to assure that the effect was not due to the amount of time spent playing with the familiar toy, nor merely to the fact that kids weren’t allowed to play with the familiar toy independently). Conversely, if the causal structure of the toy was revealed during the course of playing with it, kids always preferred the novel toy.

Finally, and perhaps most critically, the majority of children who chose to continue playing with the novel toy did so by manipulating each lever independently, thus disambiguating the toy’s causal structure!

I wish I knew more papers using such a rigorous methodology to approach exploratory play. Do any readers have recommendations?

Comments

  1. #1 Tyler Streeter
    January 2, 2008

    The papers listed below describe work on curiosity-driven robot development. In each paper the authors measure development by creating a set of increasingly complex scenarios, then plotting the amount of time spent in each scenario over time. Although the resulting behaviors are fairly simple, they do indeed progress from simple to more complex. This seems to be a fairly general way of measuring the results of “play” behavior, though it requires manually creating each situation/toy beforehand.

    It seems pretty easy to measure progress in simulations (e.g., increment a counter whenever each toy is touched). Measuring progress for real robots and humans is more difficult; you probably have to watch the subject play and record the results manually. It might be cool to write a program using computer vision that automatically tracks how long the subject plays with each toy. Then you wouldn’t have to stand there the whole time.

    Oudeyer, P.-Y., and Kaplan, F. 2004. Intelligent adaptive curiosity: a source of self-development. In Berthouze, L., et al, eds., Proceedings of the 4th International Workshop on Epigenetic Robotics, volume 117, 127-130. Lund University Cognitive Studies.

    Oudeyer, P-Y. and Kaplan, F., Hafner, V., Whyte A. (2005) The Playground Experiment: Task-Independent Development of a Curious Robot, to appear in the proceedings of the AAAI Spring Symposium Workshop on Developmental Robotics.

    Videos of the AIBO robots playing: http://playground.csl.sony.fr/en/page7.xml

    Tyler

  2. #2 CHCH
    January 2, 2008

    Hey Tyler – Thanks a lot! I’ll check it out.

    At CDS 2007 I had a nice discussion with some authors of a poster doing exactly what you describe: hand-annotated observations of play behavior with objects of varying complexity.

    The problem IMO is that there is no good definition of complexity after you’ve left information space and moved into physical reality. IIRC their definition of complexity was the number of potential ways that each toy could be used.

    This is a fine start, i guess, but it doesn’t have the same rigorous feeling as the work discussed in the post, where the hypothesis was clearly motivated, the measures unequivocal, and a series of control groups refute alternative explanations.

  3. #3 Derek James
    January 3, 2008

    Hey! Tyler…didn’t know you poked around here. Nice new website you got there.

    The problem IMO is that there is no good definition of complexity after you’ve left information space and moved into physical reality. IIRC their definition of complexity was the number of potential ways that each toy could be used.

    That sounds like a pretty lousy definition of complexity, though admittedly it’s a difficult concept. I remember enjoying playing with things like blocks and Legos as a kid. It sounds like they would consider a bag of wooden blocks more complex than a mechanical watch, which seems very misguided. In any case, what did they report? A correlation with increased play with increasing complexity or decreasing complexity? Or no correlation?

    Even though it’s more difficult to rigorously quantify, wouldn’t minimum description length as a measure of complexity still apply to physical objects?

    I just blogged about this roundtable discussion with Richard Dawkins, Lewis Wolpert, and Steve Jones about evolution and complexity. Dawkins gives a nice example of complexity by talking about the minimum level of description you would need to detail the gross anatomy of a millipede and a lobster. Because the millipede’s segments are far less differentiated than a lobster’s, much less information would be needed to specify its gross morphology than that of the lobster’s, which has fewer segments, though those segments are far more specialized and differentiated. This ignores complexity at lower levels of description (i.e. the millipede’s immune system might be far more complex than the lobster’s), but it’s a nice example.

    To be completely rigorous about it, one would need to model the physical universe to a level of detail where one could explicitly specify the amount of information needed to render a given object.

    This is veering away from your original post, but it’s a very interesting topic. To the original point, about studies involving play, I’m not familiar with any, though I had seen the study you mentioned before and thought it was very interesting. This seems related to the idea of novelty detection, that is, in this particular case if the stimulus defies their expectation it then generates more interest.

    I’ll be watching this thread with interest though. I would guess that play is a very complicated social phenomenon, and very difficult to study.

The site is currently under maintenance and will be back shortly. New comments have been disabled during this time, please check back soon.