9

I am familiar with supervised and unsupervised learning. I did the SaaS course done by Andrew Ng on Coursera.org.

I am looking for something similar for reinforcement learning.

Can you recommend something?

nbro
  • 39,006
  • 12
  • 98
  • 176
Martin S
  • 213
  • 1
  • 5

5 Answers5

7

To the good answers here, I would add

These barely scratch the surface of RL, but they should get you started.

cantordust
  • 943
  • 6
  • 10
4

Before that ask yourself if you really want to learn about "reinforcement learning." Although there is much hype about reinforcement learning, the real-world applicability of reinforcement learning is almost non-existent. Most of the online courses teach you a very little about machine learning, so it is much better to get thorough with it, rather than proceeding towards reinforcement learning. Learning reinforcement learning is somewhat different from learning about unsupervised/supervised learning techniques.

Having said that, the fastest way to get a good grasp of reinforcement learning is as follow:

  1. Read Andrej Karpathy's blog post "Pong from Pixels."

  2. Watch Deep RL Bootcamp lectures.

  3. To understand the math behind these techniques, refer to Sutton and Barto's Reinforcement Learning: An Introduction.

  4. Read relevant papers (game-playing etc.).

P.S: Make sure that you are thorough with basics of neural networks, as most of the current papers in RL involve using DNNs in some or the other way as approximators.

Yaser Sulaiman
  • 103
  • 3
thecomplexitytheorist
  • 766
  • 3
  • 19
  • `real-world applicability of reinforcement learning is almost non-existent` [AlphaGo](https://deepmind.com/research/alphago/) was trained with reinforcement learning. – cantordust Jul 03 '18 at 14:30
  • Thanks, what I like about reinforcement learning is that it can improve itself just by doing the task over and over again. No supervision is needed. Only the model has to describe the problem correctly. – Martin S Jul 03 '18 at 14:40
  • 1
    @cantordust , check https://www.alexirpan.com/2018/02/14/rl-hard.html – thecomplexitytheorist Jul 03 '18 at 15:00
  • @thecomplexitytheorist Yes, I am aware of that post, and yes, there are difficulties associated with RL (like any other ML method). However, these difficulties do not imply that its applicability is nil. Another successful example (also mentioned in the post) is the [improvement of cooling efficiency of Google's data centres by 40%](https://deepmind.com/blog/deepmind-ai-reduces-google-data-centre-cooling-bill-40/) - hardly insignificant. – cantordust Jul 03 '18 at 22:33
  • 1
    @cantordust , you should read that post completely , there is no "reinforcement learning" learning involved in "improving cooling efficiency" , they have clearly mentioned that they made predictions based on past power consumption data. – thecomplexitytheorist Jul 04 '18 at 07:17
  • @cantordust Supposed that they used RL, another problem with Google's claim is that they did not provide baselines for comparison. We can only see ML-ON and ML-OFF differences. Very simple heuristics may do much better, though this depends on the system, which we cannot look into. – Hai Nguyen Jul 04 '18 at 07:56
  • [MIT Technology review](https://www.technologyreview.com/s/601938/the-ai-that-cut-googles-energy-bill-could-soon-help-you/). Learning from past data is at the core of RL...and a lot of other ML algorithms. Also, all other things being equal, the ML on/off comparison is pretty convincing. At any rate, I am surprised at the negative reactions towards RL. It does have flaws and it is far from perfect, but it does have potential. That is precisely why we need more people working on it. – cantordust Jul 04 '18 at 11:16
4

There's a Youtube playlist (in the DeepMind channel) whose title is Introduction to reinforcement learning, which is a course (of 10 lessons) on reinforcement learning by David Silver.

A person who followed and finished the course wrote (as a Youtube comment):

Excellent course. Well paced, enough examples to provide a good intuition, and taught by someone who's leading the field in applying RL to games.

nbro
  • 39,006
  • 12
  • 98
  • 176
2

I recently saw a course by Microsoft on edx. It is called 'Reinforcement Learning Explained'.

Here is the link: https://www.edx.org/course/reinforcement-learning-explained-0 This is not quite comprehensive but at least gives a good starting point.

mash
  • 21
  • 2
1

I would say this post is a must to read:

https://rubenfiszel.github.io/posts/rl4j/2016-08-24-Reinforcement-Learning-and-DQN.html

Borislav Markov
  • 131
  • 2