1.From supervised learning to decision making problems
2.Model-free algorithms: Q-learning, policy gradients, actor-critic
3.Model-based reinforcement learning and some advanced topics and prediction
5.Transfer and multi-task learning, meta-learning
6.Open problems, research talks, invited lectures
What is reinforcement learning, and why should we care?
deep learning helps us handle unstructured environments
but doesn't tell us anyuthing about decision-making
RL gives us the mathematical framework for dealing with decision making
In RL, we have an agent that makes decisions we should call actions
the world responds with consequences we should call observations and rewards
RL actually generalizes many other machine learning
Why should we study this now
1.Advances in DL
2.Advances in RL
3.Advances in computational capability
Beyond learning from reward
Basic RL deals with maximizing rewards
This is not the only problem that matters for sequential decision making!
We will cover more advanced topics
Learing reward functions from example (inverse RL)
Transferring knowledge between domains (transfer learning, meta-learning)
Where do rewards come from?
Game --> score
well-defined notion of success might be very difficult to measure
What can DL & RL do well now?
Acquire high degree of proficiency in domains governed by simple, known rules
Learn simple skills with raw sensory inputs, given enough experience
Learn from imitating enough human-provided expert behavior
What has proven challenging so far?
Humans can learn incredibly quickly
Deep RL method are usually slow
Humans can reuse past knowledge
Transfer learning in deep RL is an open problem
Not clear what the reward function should be
Not clear what the role of prediction should be