32 min
Seminar, Hosted by Simons Institute for the Theory of Computing
Fast Reinforcement Learning With Generalized Policy Updates
McGill University & Mila / DeepMind
Recording
Abstract
Doina Precup presents a way to reduce the data demands of deep reinforcement learning by decomposing sequential decisions into tasks with their own reward functions. Generalized policy evaluation and improvement let an agent reuse previous solutions. When a new reward is close to a linear combination of known rewards, learning can reduce to linear regression. For other tasks, prior policies can still guide exploration. The work is joint with Andre Barreto, Shaobo Hou, Diana Borsa and David Silver.