I aim to build biologically-inspired reinforcement-learning agents capable of insight.
I am investigating the application of reinforcement learning with temporal abstraction in the modeling of insight in artificial agents. Insight problems require going against one's initial intuition and making use of experience to discover an original – and better – solution. In humans and animals, insight is considered a hallmark of creativity. To reproduce this in artificial agents, several challenges must be met – in particular, the combination of learning and searching, and the automatic development of temporally abstract policies through interaction with the world. An agent capable of meeting these challenges will solve insight problems in a manner consistent with the bulk of the psychological and neuroscientific literature on insight.
Nikolas Hemion; Aldebaran robotics, Paris
Tony Belpaeme, Angelo Cangelosi, Michaela Gummerum, Thomas Wennekers (Plymouth University), Nikolas Hemion (Aldebaran robotics)