Towards Fast Autonomous Learners: Advances in Reinforcement Learning - 2015

Explore cutting-edge research on creating fast autonomous learners in this 57-minute lecture by Emma Brunskill at the Center for Language & Speech Processing, JHU. Delve into the challenges of developing AI agents that can make good decisions in stochastic environments, with a focus on applications involving human interaction. Learn about transfer learning across sequential decision-making tasks and its potential to improve educational tools. Discover key concepts such as Markov Decision Processes, reinforcement learning, and unbiased policy evaluation. Examine innovative approaches like queue-based offline evaluation and Bayesian optimization for faster, more efficient policy search. Investigate personalization and transfer learning techniques for sequential decision-making tasks, and understand the importance of sample efficiency in real-world applications. Gain insights from Brunskill's expertise in AI and machine learning, particularly in the context of intelligent tutoring systems and medical interfaces.

Intro
Markov Decision Process (MDP)
Reinforcement Learning
Unbiased Policy Evaluation for General RL in Short Horizons
Queue-based Offline Evaluation of Online Bandit Algorithms
Our Queue Approach Can Sometimes Evaluate Algorithms that Use Deterministic Policies for Many More Time Steps than Rejection
Sample Complexity of RL
Provably More Efficient Learners
Fast, Better Policy Search using Bayesian Optimization
Black Box Optimization
Opening the Box: Leverage Offline Policy Evaluation
Personalization & Transfer Learning for Sequential Decision Making Tasks
Latent Variable Modeling Background
Diameter Assumption: Needed for Sample Complexity Improvement in Transfer?
Active Set is Models Compatible with Current Task's Data
More Data Efficient Learning In Domains Where It Matters