Thomas R. Colin


Insight is perhaps the cognitive phenomenon most closely associated with creativity. People engaged in problem-solving sometimes experience a sudden transformation: they see the problem in a radically different manner, and simultaneously feel with great certainty that they have found the right solution. The change of problem representation is called "restructuring", and the affective changes associated with sudden progress are called the "Aha!" experience. Together, restructuring and the "Aha!" experience characterize insight. Reinforcement Learning is both a theory of biological learning and a subfield of machine learning. In its psychological and neuroscientific guise, it is used to model habit formation, and, increasingly, executive function. In its artificial intelligence guise, it is currently the favored paradigm for modeling agents interacting with an environment. Reinforcement learning, I argue, can serve as a model of insight: its foundation in learning coincides with the role of experience in insight problem-solving; its use of an explicit "value" provides the basis for the "Aha!" experience; and finally, in a hierarchical form, it can achieve a sudden change of representation resembling restructuring. An experiment helps confirm some parallels between reinforcement learning and insight. It shows how transfer from prior tasks results in considerably accelerated learning, and how the value function increase resembles the sense of progress corresponding to the "Aha!"-moment. However, a model of insight on the basis of hierarchical reinforcement learning did not display the expected "insightful" behavior. A second model of insight is presented, in which temporal abstraction is based on self-prediction: by predicting its own future decisions, an agent adjusts its course of action on the basis of unexpected events. This kind of temporal abstraction, I argue, corresponds to what we call "intentions", and offers a promising model for biological insight. It explains the "Aha!" experience as resulting from a temporal difference error, whereas restructuring results from an adjustment of the agent's internal state on the basis of either new information or a stochastic interpretation of stimuli. The model is called the actor-critic-intention (ACI) architecture. Finally, the relationship between intentions, insight, and creativity is extensively discussed in light of these models: other works in the philosophical and scientific literature are related to, and sometimes illuminated by the ACI architecture.

Document Type


Publication Date