Reinforcement Learning, not Supervised Learning, Can Lead to Insight
- Arata Nonami, The University of Tokyo, Tokyo, Japan
- Haruaki Fukuda, The University of Tokyo, Tokyo, Japan
- Yoshiyuki Sato, University of Tokyo, Tokyo, Japan
- Kazuyuki Samejima, Tamagawa University, Tokyo, Japan
- Kazuhiro Ueda, The University of Tokyo, Tokyo, Japan
AbstractThis study examined the differences among individuals in the performance of insight problem solving. The problem-solving characteristics of an individual seemed to be dependent on what and how they had learned. Thus, we compared the performances of insight problem solving between reinforcement and supervised learners. The results showed that the performances of reinforcement learners were better than those of supervised learners, although the non-insight problem solving performance of both learner types was comparable. This result suggests that insight might be supported by the cognitive mechanisms underlying reinforcement learning. In particular, we speculate that the degree of exploration, by which reinforcement learning is characterized, might have an impact on the performance of insight problem solving