Daydreaming Algorithms

The project, funded by the Villum Foundation through the Villum Experiment programme, aims at enabling algorithms to mimic the human ability to daydream, i.e. to enable reinforcement learning algorithms to train without direct sensory or reward input using learning models of the world. By coupling recent neuroscience theories with the latest developments in reinforcement learning, we aim at developing algorithms that train with less interaction with the environment in which they act, and which are able to generalize trained behavior over multiple tasks.

The project started October 2019 and runs until late 2021.