Lecture 3: Interaction in Imitation Learning скачать в хорошем качестве

Lecture 3: Interaction in Imitation Learning

In this third lecture, we dive to the core of imitation learning to understand the role of interaction. Unlike traditional supervised learning where one gets data upfront, robots have to interact with both humans (imitation learning) and the world (system identification) to acquire such data. What makes for a good interaction strategy? We take a look at no-regret online learning for insights and derive practical yet powerful algorithms for interaction. For more information about me and my work, check out http://www.sanjibanchoudhury.com/ References: 1. Ross et al. "A Reduction of Imitation Learning and Structured Prediction to No-Regret Online Learning" https://arxiv.org/abs/1011.0686 2. Ross and Bagnell. "Agnostic System Identification for Model-Based Reinforcement Learning" https://arxiv.org/abs/1203.1007 3. Kakade and Shai Shalev-Schwartz. "Mind the Duality Gap: Logarithmic regret algorithms for online optimization" https://papers.nips.cc/paper/2008/has... 4. Cheng and Boots. "Convergence of Value Aggregation for Imitation Learning" http://proceedings.mlr.press/v84/chen...