Expected information gain predicts curiosity

Abstract: Curiosity has been defined as an intrinsic motivation for performing actions that result in a gain of knowledge (Berlyne, 1966; Loewenstein, 1994). As positive effects of curiosity on memorization of new knowledge were found by Kang et al., 2009, exploring the mechanisms behind curiosity can have great practical applications. Although information theoretic concepts have been linked to curiosity (Berlyne, 1966; Kang et al., 2009; Gottlieb, Oudeyer, Lopes & Baranes, 2013), this has not been tested empirically through an actual information theoretic operationalisation. This study aims to correct this by individually measuring the prior knowledge of participants and computing the expected information gain (IG) of an information rewarding action. To quantify the incentive value of curiosity a time delay was imposed on IG, forcing participants to a trade-off between maximizing information and minimizing time spent on the task. Through linear regression analysis, it is shown that curiosity is proportional to expected IG and that participants were prepared to invest 0.4 seconds per bit.