/Practical_RL

A course in reinforcement learning in the wild

Primary LanguageJupyter NotebookMIT LicenseMIT

Practical_RL

A course on reinforcement learning in the wild. Taught on-campus in HSE and Yandex SDA (russian) and maintained to be friendly to online students (both english and russian).

Manifesto:

  • Optimize for the curious. For all the materials that aren’t covered in detail there are links to more information and related materials (D.Silver/Sutton/blogs/whatever). Assignments will have bonus sections if you want to dig deeper.
  • Practicality first. Everything essential to solving reinforcement learning problems is worth mentioning. We won't shun away from covering tricks and heuristics. For every major idea there should be a lab that allows to “feel” it on a practical problem.
  • Git-course. Know a way to make the course better? Noticed a typo in a formula? Found a useful link? Made the code more readable? Made a version for alternative framework? You're awesome! Pull-request it!

Coordinates and useful links

Announcements

  • 25.03.16 - HSE important next monday lecture is postponed by 1 week due to HSE mid-term exams. Deadlines have been postponed accordingly.
  • 25.03.16 - week5 you can submit any atari game you want.
  • 16.03.16 - week4 homework max score threshold for LunarLander reduced to -100
  • 16.03.16 - (hse) shifted deadline for week5
  • 15.03.16 - (hse) added week6 assignment and deadline
  • 10.03.17 - (ysda/hse students) important please consider Course Projects as an alternative way of completing the course.
  • 8.03.17 - YSDA deadlines announced for weeks 3 and 3.5, sry for only doing this now.
  • 01.03.17 - YSDA deadline on week2 homework moved to 08.03.17
Previous announcements

* 28.02.17 - (HSE) homework 4 published * 24.02.17 - Dependencies updated ([same url](yandexdataschool#1)). Please install theano/lasagne/agentnet until week4 or make sure you're familiar enough with your deep learning framework of choice. * 23.02.17 - YSDA homework 2 can be found [here](https://github.com/yandexdataschool/Practical_RL/tree/master/week2). If you're from HSE you can opt to submit either old or new whichever you prefer. * 17.02.17 - warning! we force-pushed into the repository. Please back-up your github files before you pull! * 16.02.17 - Lecture slides are now available through urls in README files for each week like [this](https://github.com/yandexdataschool/Practical_RL/tree/master/week1#materialshttps://github.com/yandexdataschool/Practical_RL/tree/master/week1#materials). You can also find full archive [here](https://yadi.sk/d/loPpY45J3EAYfU). * 30.03.17 - YSDA deadlines announced for HW 4 * 16.02.17 - HSE homework 3 added * 14.02.17 - HSE deadlines for weeks 1-2 extended! * 14.02.17 - anytask invites moved [here](https://github.com/yandexdataschool/Practical_RL/wiki/Homeworks-and-grading-(HSE-and-YSDA)) * 14.02.17 - if you're from HSE track and we didn't reply to your week0 homework submission, raise panic! * 11.02.17 - week2 success thresholds are now easier: get >+50 for LunarLander or >-180 for MountainCar. Solving env will yield bonus points. * 13.02.17 - Added invites for anytask.org * 10.02.17 - from now on, we'll formally describe homework and add useful links via ./week*/README.md files. [Example.](https://github.com/yandexdataschool/Practical_RL/blob/master/week0/README.md) * 9.02.17 - YSDA track started * 7.02.17 - HWs checked up * 6.02.17 - week2 uploaded * 27.01.17 - merged fix by _omtcyfz_, thanks! * 27.01.17 - added course mail for homework submission: __practicalrl17@gmail.com__ * 23.01.17 - first class happened * 23.01.17 - created repo

Syllabus

  • week0 Welcome to the MDP

    • Lecture: RL problems around us. Markov decision process. Simple solutions through combinatoric optimization.
    • Seminar: Frozenlake with genetic algorithms
    • Homework description - week0/README.md
      • HSE Homework deadline: 23.59 1.02.17
      • YSDA Homework deadline: 23.59 19.02.17
  • week1 Crossentropy method and monte-carlo algorithms

    • Lecture: Crossentropy method in general and for RL. Extension to continuous state & action space. Limitations.
    • Seminar: Tabular CEM for Taxi-v0, deep CEM for box2d environments.
    • Homework description - week1/README.md
      • HSE homework deadline: 23.59 15.02.17
      • YSDA homework deadline: 23.59 26.02.17
  • week2 Temporal Difference

    • Lecture: Discounted reward MDP. Value iteration. Q-learning. Temporal difference Vs Monte-Carlo.
    • Seminar: Tabular q-learning
    • Homework description - week2/README.md
      • HSE homework deadline: 23.59 15.02.17
      • YSDA homework deadline: 23.59 8.03.17
  • week3 Value-based algorithms

    • Lecture: SARSA. Off-policy Vs on-policy algorithms. N-step algorithms. Eligibility traces.
    • Seminar: Qlearning Vs SARSA Vs expected value sarsa in the wild
    • Homework description - week3/README.md
      • HSE homework deadline 23.59 22.02.17
      • YSDA homework deadline: 23.59 14.03.17
  • week3.5 Deep learning recap

    • Lecture: deep learning, convolutional nets, batchnorm, dropout, data augmentation and all that stuff.
    • Seminar: Theano/Lasagne on mnist, simple deep q-learning with CartPole (TF version contrib is welcome)
    • Homework - convnets on MNIST or simple deep q-learning - week3.5/README.md
      • HSE homework deadline 23.59 1.03.17
      • YSDA homework deadline: 23.59 14.03.17 (5 pts)
  • week4 Approximate reinforcement learning

    • Lecture: Infinite/continuous state space. Value function approximation. Convergence conditions. Multiple agents trick.
    • Seminar: Approximate Q-learning with experience replay. (CartPole, Acrobot, Doom)
    • Homework - q-learning manually, experience replay - week4/README.md
      • HSE homework deadline 23.59 8.03.17
      • YSDA homework deadline 23.59 19.03.17
  • week5 Deep reinforcement learning

    • Lecture: Deep Q-learning/sarsa/whatever. Heuristics & motivation behind them: experience replay, target networks, double/dueling/bootstrap DQN, etc.
    • Seminar: DQN on atari
    • Homework - Breakout with DQN and advanced tricks - week5/README.md
      • HSE homework deadline 23.59 22.03.17
      • YSDA homework deadline 23.59 26.03.17
  • week6 Policy gradient methods

    • Lecture: Motivation for policy-based, policy gradient, logderivative trick, REINFORCE/crossentropy method, variance theorem(advantage), advantage actor-critic (incl.n-step advantage)
    • Seminar: REINFORCE manually, advantage actor-critic for MountainCar - week6/README.md
      • HSE homework deadline 23.59 2.04.17
      • YSDA deadline 23.59 6.04.2017
  • week6.5 RNN recap

    • Lecture: recurrent neura networks for sequences. GRU/LSTM. Gradient clipping. Seq2seq
    • Seminar: char-rnn and simple seq2seq
      • HSE homework deadline 23.59 5.04.17
      • YSDA deadline 23.59 9.04.2017
  • week7 Partially observable MDPs

    • Lecture: POMDP intro. Model-based solvers. RNN solvers. RNN tricks: attention, problems with normalization methods, pre-training.
    • Seminar: Deep kung-fu & doom with recurrent A3C and DRQN
      • HSE homework deadline 23.59 16.04.17 (first submission; kung fu assignment is worth 6pts isntead of 3)
      • YSDA homework deadline 23.59 19.04.17 (first submission)
  • week 8 Case studies 1 (approx 10.04.17)

    • Lecture: Reinforcement Learning as a general way to optimize non-differentiable loss. Seq2seq tasks: g2p, machine translation, conversation models, image captioning.
    • Seminar: Simple neural machine translation with self-critical policy gradient
      • HSE deadline 23.59 30.04.17 (first submission)
      • YSDA deadline 23.59 30.05.17 (first submission)

Future lectures:

  • week 9 RL in Large/Continuous action spaces. (approx 17.04.17)

    • Lecture: Continuous action space MDPs. Value-based approach (NAF). Special case algorithms (dpg, svg). Case study:finance. Large discrete action space problem. Action embedding.
    • Seminar: Classic Control and BipedalWalker with ddpg Vs qNAF. https://gym.openai.com/envs/BipedalWalker-v2 . Financial bot as bonus track.
  • week 10 Trust Region Policy Optimization. (approx 24.05.17)

    • Lecture: Trust region policy optimization in detail. NPO/TRPO.
    • Seminar: approximate TRPO vs approximate Q-learning for gym box2d envs (robotics-themed).
  • week 11 Advanced exploration methods (appox 1.05.17)

    • Lecture1: Improved exploration methods (quantile-based, etc.) for bandits. Bayesian approach. Case study: Contextual bandits for RTB.
    • Lecture2: Augmented rewards. Density-based models, UNREAL, variational information maximizing exploration, bayesian optimization with BNNs.
  • week 12 Advanced RL topics (approx 8.05.17)

    • Lecture 1: Hierarchical MDP. MDP Vs real world. Sparse and delayed rewards. When Q-learning fails. Hierarchical MDP. Hierarchy as temporal abstraction. MDP with symbolic reasoning.
  • Lecture 2: Knowledge Transfer in RL & Inverse Reinforcement Learning: basics; personalized medical treatment; robotics.

Course staff

Course materials and teaching by

Contributors