Daydreaming Algorithms

Can algorithms daydream? Computer science, specifically the subfield machine learning, has the aim of developing algorithms that can learn to perform tasks without human guidance. Reinforcement learning algorithms now often exhibit super-human performance at playing, for example, simple computer games. But the algorithms use many observations: They often play millions of games during their training. The many repetitions often prevent practical use of reinforcement learning algorithms beyond game playing. In neuroscience, current hypotheses state that the human brain has an internal abstract model of the world as it tests against sensory input. The brain can use this model to train without observing millions of games directly. Instead, humans 'daydream'. In the project, we will investigate whether reinforcement learning algorithms can daydream and thereby learn to perform tasks with much fewer observations.