Umeå universitets logga

The aim of this thesis is to investigate the computational mechanisms that drive curiosity, the motivational state for learning for its own sake. Curiosity is operationalized as information-seeking behaviors and self-reported curiosity. A novel mechanism based on resolving uncertainty in episodic events (episodic curiosity) is tested and compared to predominant theories on curiosity. This computational mechanism is then used to investigate how curiosity interacts with other types of motivation, and the effects of curiosity on learning in different knowledge domains. Specifically, individually measured indicators of uncertainty are computed using an information-theoretic quantity of entropy (Shannon, 1948).

The aim of the first study is to compare an existing theory, the Learning Progress hypothesis (Gottlieb & Oudeyer, 2018; Oudeyer et al., 2016), to episodic curiosity account. The experiment takes the form of an asteroid avoidance computer game, where outcome uncertainty can be dispelled by requesting trial-related feedback. The results of this study indicate that a desire to close episodic uncertainty is enough to drive and maintain information-seeking behavior, but learning progress is also a lesser determinant of information-seeking.

In the second study the relationship between uncertainty, curiosity and satisfaction is examined. The experiment tests a predominant theory that curiosity reflects a desire for pleasurable learning. The effect of uncertainty and curiosity on memory encoding is also examined. The results show that satisfaction had a negative association with learning, indicating that the learning resulting from curiosity may not necessarily be pleasurable.

The final study looks at how curiosity integrates with other motivational goals, by aligning curiosity with or against the best option in a two-armed bandit task. Can curiosity distract from, or improve, learning and performance in the bandit task? This study shows that the motivational value from curiosity can integrate with extrinsic task goals to improve or distract from performance.