Proceedings of the Annual Meeting of the Cognitive Science Society

A few minutes practice is often more than sufficient for an adult human participant to identify the structure of a problem they have never seen before, and plan a complex sequence of actions that creates a solution. However, it remains less clear whether this distinctive ability for ad-hoc discovery of problem-solving algorithms is itself subject to rapid meta-learning. We developed a novel problem-solving paradigm to examine aspects of this question. Participants in our study faced repeated iterations of an interactive sequential reasoning problem. Over trials, those who faced the hardest version developed increasingly efficient hierarchically-structured strategies that adaptively sequence a subordinate learning algorithm and an action planning policy; those who faced an easier version used simpler action-based strategies that did not involve learning the underlying structure. These results offer experimental evidence for efficient meta-learning of algorithmic concepts in a problem-solving setting.