Repository files navigation

• Simple guide and collective to study RL/DeepRL in one to 2.5 months of time.

• Introduction to Reinforcement Learning by Joelle Pineau, McGill University:

  • Applications of RL.

  • When to use RL?

  • RL vs supervised learning

  • What is MDP? Markov Decision Process

  • Components of an RL agent:

    • states
    • actions (Probabilistic effects)
    • Reward function
    • Initial state distribution

                                    +-----------------+       +--------------------- |                 |       |                      |      Agent      |       |                      |                 | +---------------------+       |         +----------> |                 |                       |       |         |            +-----------------+                       |       |         |                                                      | state |         | reward                                               | action S(t)  |         | r(t)                                                 | a(t)       |         |                                                      |       |         | +                                                    |       |         | |  r(t+1) +----------------------------+             |       |         +-----------+                            |             |       |           |         |                            | <-----------+       |           |         |      Environment           |       |           |  S(t+1) |                            |       +---------------------+                            |                   |         +----------------------------+                   + * Sutton and Barto (1998)

  • Explanation of the Markov Property:

  • Why Maximizing utility in:

    • Episodic tasks
    • Continuing tasks
      • The discount factor, gamma γ

  • What is the policy & what to do with it?

    • A policy defines the action-selection strategy at every state:

  • Value functions:

    • The value of a policy equations are (two forms of) Bellman’s equation.
    • (This is a dynamic programming algorithm).
    • Iterative Policy Evaluation:
      • Main idea: turn Bellman equations into update rules.

  • Optimal policies and optimal value functions.

    • Finding a good policy: Policy Iteration (Check the talk Below By Peter Abeel)
    • Finding a good policy: Value iteration
      • Asynchronous value iteration:
      • Instead of updating all states on every iteration, focus on important states.

  • Key challenges in RL:

    • Designing the problem domain
      • State representation– Action choice– Cost/reward signal
    • Acquiring data for training– Exploration / exploitation– High cost actions– Time-delayed cost/reward signal
    • Function approximation
    • Validation / confidence measures

  • The RL lingo.

  • In large state spaces: Need approximation:

    • Fitted Q-iteration:
      • Use supervised learning to estimate the Q-function from a batch of training data:
      • Input, Output and Loss.
        • i.e: The Arcade Learning Environment

  • Deep Q-network (DQN) and tips.

• Deep Reinforcement Learning by Pieter Abbeel, EE & CS, UC Berkeley

  • Why Policy Optimization?

  • Cross Entropy Method (CEM) / Finite Differences / Fixing Random Seed

  • Likelihood Ratio (LR) Policy Gradient

  • Natural Gradient / Trust Regions (-> TRPO)

  • Actor-Critic (-> GAE, A3C)

  • Path Derivatives (PD) (-> DPG, DDPG, SVG)

  • Stochastic Computation Graphs (generalizes LR / PD)

  • Guided Policy Search (GPS)

  • Inverse Reinforcement Learning

    • Inverse RL vs. behavioral cloning

  • Explanation with Implementation for some of the topics mentioned in the Deep Reinforcement Learning talk, written byArthur Juliani

    • The TF / Python implementationscan be found here.
    • Part 0 — Q-Learning Agents
    • Part 1 — Two-Armed Bandit
    • Part 1.5 — Contextual Bandits
    • Part 2 — Policy-Based Agents
    • Part 3 — Model-Based RL
    • Part 4 — Deep Q-Networks and Beyond
    • Part 5 — Visualizing an Agent’s Thoughts and Actions
    • Part 6 — Partial Observability and Deep Recurrent Q-Networks
    • Part 7 — Action-Selection Strategies for Exploration
    • Part 8 — Asynchronous Actor-Critic Agents (A3C)

• Before starting out the books, here is a neat overview by Yuxi Li about Deep RL:
  • Deep Reinforcement Learning: An Overview
• Reinforcement Learning: An Introduction by Richard S. Sutton and Andrew G. Barto
• Algorithms for Reinforcement Learning.
• Reinforcement Learning and Dynamic Programming using Function Approximators.

• Reinforcement Learning by David Silver.

  • Lecture 1: Introduction to Reinforcement Learning
  • Lecture 2: Markov Decision Processes
  • Lecture 3: Planning by Dynamic Programming
  • Lecture 4: Model-Free Prediction
  • Lecture 5: Model-Free Control
  • Lecture 6: Value Function Approximation
  • Lecture 7: Policy Gradient Methods
  • Lecture 8: Integrating Learning and Planning
  • Lecture 9: Exploration and Exploitation
  • Lecture 10: Case Study: RL in Classic Games

• CS 294: Deep Reinforcement Learning, Spring 2017 by John Schulman and Pieter Abbeel.

  • Instructors: Sergey Levine, John Schulman, Chelsea Finn:
  • My Bad Notes