Grokking Deep Reinforcement Learning

Note: At the moment, only running the code from the docker container (below) is supported. Docker allows for creating a single environment that is more likely to work on all systems. Basically, I install and configure all packages for you, except docker itself, and you just run the code on a tested environment.

To install docker, I recommend a web search for "installing docker on <your os here>". For running the code on a GPU, you have to additionally install nvidia-docker. NVIDIA Docker allows for using a host's GPUs inside docker containers. After you have docker (and nvidia-docker if using a GPU) installed, follow the three steps below.

Running the code

0. Clone this repo:
   git clone --depth 1 https://github.com/mimoralea/gdrl.git && cd gdrl
1. Pull the gdrl image with:
   docker pull mimoralea/gdrl:v0.14
2. Spin up a container:
   • On Mac or Linux:
     docker run -it --rm -p 8888:8888 -v "$PWD"/notebooks/:/mnt/notebooks/ mimoralea/gdrl:v0.14
   • On Windows:
     docker run -it --rm -p 8888:8888 -v %CD%/notebooks/:/mnt/notebooks/ mimoralea/gdrl:v0.14
   • NOTE: Use nvidia-docker if you are using a GPU.
3. Open a browser and go to the URL shown in the terminal (likely to be: http://localhost:8888). The password is: gdrl

About the book

Book's website

https://www.manning.com/books/grokking-deep-reinforcement-learning

Table of content

1. Introduction to deep reinforcement learning
2. Mathematical foundations of reinforcement learning
3. Balancing immediate and long-term goals
4. Balancing the gathering and utilization of information
5. Evaluating agents' behaviors
6. Improving agents' behaviors
7. Achieving goals more effectively and efficiently
8. Introduction to value-based deep reinforcement learning
9. More stable value-based methods
10. Sample-efficient value-based methods
11. Policy-gradient and actor-critic methods
12. Advanced actor-critic methods
13. Towards artificial general intelligence

Detailed table of content

1. Introduction to deep reinforcement learning

• (Livebook)
• (No Notebook)

2. Mathematical foundations of reinforcement learning

• (Livebook)
• (Notebook)
  • Implementations of several MDPs:
    • Bandit Walk
    • Bandit Slippery Walk
    • Slippery Walk Three
    • Random Walk
    • Russell and Norvig's Gridworld from AIMA
    • FrozenLake
    • FrozenLake8x8

3. Balancing immediate and long-term goals

• (Livebook)
• (Notebook)
  • Implementations of methods for finding optimal policies:
    • Policy Evaluation
    • Policy Improvement
    • Policy Iteration
    • Value Iteration

4. Balancing the gathering and utilization of information

• (Livebook)
• (Notebook)
  • Implementations of exploration strategies for bandit problems:
    • Random
    • Greedy
    • E-greedy
    • E-greedy with linearly decaying epsilon
    • E-greedy with exponentially decaying epsilon
    • Optimistic initialization
    • SoftMax
    • Upper Confidence Bound
    • Bayesian

5. Evaluating agents' behaviors

• (Livebook)
• (Notebook)
  • Implementation of algorithms that solve the prediction problem (policy estimation):
    • On-policy first-visit Monte-Carlo prediction
    • On-policy every-visit Monte-Carlo prediction
    • Temporal-Difference prediction (TD)
    • n-step Temporal-Difference prediction (n-step TD)
    • TD(λ)

6. Improving agents' behaviors

• (Livebook)
• (Notebook)
  • Implementation of algorithms that solve the control problem (policy improvement):
    • On-policy first-visit Monte-Carlo control
    • On-policy every-visit Monte-Carlo control
    • On-policy TD control: SARSA
    • Off-policy TD control: Q-Learning
    • Double Q-Learning

7. Achieving goals more effectively and efficiently

• (Livebook)
• (Notebook)
  • Implementation of more effective and efficient reinforcement learning algorithms:
    • SARSA(λ) with replacing traces
    • SARSA(λ) with accumulating traces
    • Q(λ) with replacing traces
    • Q(λ) with accumulating traces
    • Dyna-Q
    • Trajectory Sampling

8. Introduction to value-based deep reinforcement learning

• (Livebook)
• (Notebook)
  • Implementation of a value-based deep reinforcement learning baseline:
    • Neural Fitted Q-iteration (NFQ)

9. More stable value-based methods

• (Livebook)
• (Notebook)
  • Implementation of "classic" value-based deep reinforcement learning methods:
    • Deep Q-Networks (DQN)
    • Double Deep Q-Networks (DDQN)

10. Sample-efficient value-based methods

• (Livebook)
• (Notebook)
  • Implementation of main improvements for value-based deep reinforcement learning methods:
    • Dueling Deep Q-Networks (Dueling DQN)
    • Prioritized Experience Replay (PER)

11. Policy-gradient and actor-critic methods

• (Livebook)
• (Notebook)
  • Implementation of classic policy-based and actor-critic deep reinforcement learning methods:
    • Policy Gradients without value function and Monte-Carlo returns (REINFORCE)
    • Policy Gradients with value function baseline trained with Monte-Carlo returns (VPG)
    • Asynchronous Advantage Actor-Critic (A3C)
    • Generalized Advantage Estimation (GAE)
    • [Synchronous] Advantage Actor-Critic (A2C)

12. Advanced actor-critic methods

• (Livebook)
• (Notebook)
  • Implementation of advanced actor-critic methods:
    • Deep Deterministic Policy Gradient (DDPG)
    • Twin Delayed Deep Deterministic Policy Gradient (TD3)
    • Soft Actor-Critic (SAC)
    • Proximal Policy Optimization (PPO)

13. Towards artificial general intelligence

• (Livebook)
• (No Notebook)