Reinforcement Learning Upside Down - Don't Predict Rewards, Just Map Them to Actions

Explore a groundbreaking approach to reinforcement learning in this 27-minute video by Yannic Kilcher. Dive into Juergen Schmidhuber's innovative "Upside-Down Reinforcement Learning" (UDRL) concept, which transforms traditional reinforcement learning into a form of supervised learning. Learn how UDRL constructs a behavior function that uses desired rewards as input, rather than predicting them. Discover the surprising performance of this new paradigm compared to classic RL algorithms. Understand how UDRL interprets input observations as commands, mapping them to actions through supervised learning on past experiences. Examine the potential applications of UDRL in achieving high rewards and other goals within specified time horizons. Additionally, explore a related approach for teaching robots to imitate humans through the "Imitate-Imitator" concept. Gain insights into how this innovative thinking may explain certain aspects of biological evolution and parent-child interactions.

Reinforcement Learning Upside Down: Don't Predict Rewards -- Just Map Them to Actions