Core Concepts: Interactive No-Regret Learning скачать в хорошем качестве

Core Concepts: Interactive No-Regret Learning

We explore the concept of interactive learning. Robots must interact with the world to gather data on which they learn. The principled way to learn when your data can be changing, possibly adversarially, is by striving to be "no regret", i.e., do as well as the best policy in hindsight. But greedily picking the best policy in hindsight fails, even on the simplest of examples! Join us as we journey through the world of games to understand why and how the simple act of hedging not only achieves "no regret", but unlocks some of the most powerful algorithms in the universe! We acknowledge Drew Bagnell for many insightful conversations on this topic. Check out the full series "Core Concept in Robotics":    • Core Concepts in Robotics   For a deeper dive, check out the series "Imitation Learning: A Series of Deep Dives":    • Imitation Learning: A Series of Deep Dives   References: 1. Drew Bagnell lecture notes: http://www.cs.cmu.edu/~16831-f14/note... 2. Blum et al. "On-Line Algorithms in Machine Learning" https://www.cs.cmu.edu/~ninamf/ML10/o... 3. Shai Shalev-Shwartz et al. "Online Learning and Online Convex Optimization" https://www.cs.huji.ac.il/w~shais/pap... 4. Arora et al. "The Multiplicative Weights Update Method: a Meta Algorithm and Applications" https://www.cs.princeton.edu/~arora/p... 5. Kakade et al. "Mind the Duality Gap: Logarithmic regret algorithms for online optimization" https://proceedings.neurips.cc/paper/...