The complexity of model-free and model-based learning strategies
Alexandre L. S. Filipowicz, Jonathan Levine, Eugenio Piasini, Gaia Tavoni, Joseph W. Kable, Joshua I. Gold
Received Date: 8th January 20
A proposed continuum of learning strategies, from model-free to model-based, is thought to progress systematically in complexity and therefore flexibility. Here we distinguish different forms of complexity to show that, contrary to this idea, strategies at both ends of this continuum can be equally flexible. Using a canonical learning task, we first simulated behavior to show that computational complexity, a measure of implementation demands, is higher for a standard model-based versus model-free algorithm, but information complexity, a measure of flexibility, is not. We then analyzed human behavior to show that information complexity, which unlike computational complexity can be estimated from behavior, tended to increase for strategies that were increasingly either model-free or model-based, resulting in similar accuracy, suboptimal use of information, and increased response times. Thus, model-free and model-based strategies can have similar overall flexibility and instead are better distinguished by the specific task features from which they learn.
Read in full at bioRxiv.
This is an abstract of a preprint hosted on an independent third party site. It has not been peer reviewed but is currently under consideration at Nature Communications.