/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? Made the code more readable? Made a version for alternative framework? You're awesome! Pull-request it!

Coordinates and useful links

Announcements

  • 16.02.17 - Lectures moved
  • 16.02.17 - HSE homework 3 added
  • 14.02.17 - HSE deadlines for weeks 1-2 extended!
  • 14.02.17 - anytask invites moved here
  • 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.
  • 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 Monte-carlo methods
  • 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.
    • HSE homework deadline: 23.59 15.02.17
  • week2 Temporal Difference
  • Lecture: Discounted reward MDP. Value iteration. Q-learning. Temporal difference Vs Monte-Carlo.
  • Seminar: Tabular q-learning
  • 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
    • HSE homework deadline 23.59 22.02.17

Future lectures:

  • week4 Approximate reinforcement learning
  • Lecture: Infinite/continuous state space. Value function approximation. Convergence conditions. Multiple agents trick.
  • Seminar: Approximate Q-learning. (CartPole, MountainCar, Breakout)

somewhere here introduction to theano

  • week i+1 Deep reinforcement learning

  • Lecture: Deep Q-learning/sarsa/whatever. Heuristics & motivation behind them: experience replay, target networks, double/dueling/bootstrap DQN, etc.

  • Seminar: Playing atari with deep reinforcement learning. Experience replay. (classwork = doombasic)

  • week i+1 Policy-based methods

  • Lecture: Motivation for policy-based, policy gradient, logderivative trick, REINFORCE/crossentropy method, variance theorem(advantage), advantage actor-critic (incl.n-step advantage), off-policy actor-critic (off-PAC), natural gradients(briefly), continuous action space(teaser).

  • Seminar: a2c Vs qlearning for MountainCar/Doom, entropy regularization & tricks.

  • week i+1 Trust Region Policy Optimization.

  • Lecture: Trust region policy optimization in detail.

  • approximate TRPO vs approximate Q-learning for gym box2d envs (robotics-themed)

  • week i+1 Large/Continuous action space. Case study: recsys.

  • Lecture: Continuous action space MDPs. Model-based approach (NAF). Actor-critic approach (dpg, svg). Trust Region Policy Optimization. Large discrete action space problem. Action embedding.

  • Seminar: Classic Control and BipedalWalker with ddpg Vs qNAF. https://gym.openai.com/envs/BipedalWalker-v2 .

somewhere here RNN crash-course

  • week i+1 Partially observable MDPs

  • Lecture: POMDP intro. Model-based solvers. RNN solvers. RNN tricks: attention, problems with normalization methods, pre-training.

  • Seminar: Deep kung-fu with recurrent A2C vs feedforward A2C

  • week i+1 Advanced exploration methods: intrinsic motivation

  • Lecture: Augmented rewards. Heuristics (UNREAL,density-based models), formal approach: information maximizing exploration. Model-based tricks(also refer mcts).

  • Seminar: Vime vs epsilon-greedy for Go9x9 (bonus 19x19)

  • week i+1 Advanced exploration methods: probablistic approach.

  • Lecture: Improved exploration methods (quantile-based, etc.). Bayesian approach. Case study: Contextual bandits for RTB.

  • Seminar: Bandits

  • week i+1 Case studies I

  • Lecture: Reinforcement Learning as a general way to optimize non-differentiable loss. KL(p||q) vs KL(q||p). Case study: machine ranslation, speech synthesis, conversation models.

  • Seminar: Optimizing Levenstein for word transcription

  • week i+1 Hierarchical MDP

  • Lecture: MDP Vs real world. Sparse and delayed rewards. When Q-learning fails. Hierarchical MDP. Hierarchy as temporal abstraction. MDP with symbolic reasoning.

  • Seminar: Hierarchical RL for atari games with rare rewards (starting from pre-trained DQN)

  • week i+1 Case studies II

  • Lecture: Direct policy optimization: finance. Inverse Reinforcement Learning: personalized medial treatment, robotics.

  • Seminar: Portfolio optimization as POMDP.

Course staff

Course materials and teaching by

Contributors