Reinforcement Learning

Reinforcement Learning

author

Adam Kolář

Prerequisites

  • basic knowledge of programing in Python, PyTorch, Numpy, Pandas
  • high school level of mathematics
  • Basics of machine learning on the level of our course Introduction to machine Learning

Abstract

Reinforcement learning is one of the most progressive parts of machine learning these days. Great breakthroughs in the field of playing games helped a lot to popularize and improve methods dealing with understanding sparse information distributed across long time gaps.

One of the great examples of those successes is for instance AlphaZero algorithm used to train models to play e.g. the game of Go. Those models were able to defeat various human grandmasters, something we expected to happen (maybe) in the upcoming decades.

In the workshop, we will explore basic concepts of reinforcement learning on discrete problem space. This will help us to introduce neural networks as models for value and policy based methods, which we will apply to several popular game environments. The same approaches are incorporated in the state of the art solutions, which we will discuss in the last part of the workshop.

Outline

  • Introduction to PyTorch & OpenAI gym
  • Introduction to reinforcement learning
    • State value functions
    • Action value functions
    • Monte carlo approach
  • Value based methods
    • Temporal difference methods
    • Q-learning
    • Replay buffers
    • Hands on application to some of the game environments
  • Policy based methods
    • The reinforce algorithm
    • Credit assignment improvement
    • Hands on application to some of the game environments
  • Actor critic methods
    • Basic setup explanation
    • DDPG algorithm
    • Hands on application to some of the game environments
  • AlphaZero
    • Basic concepts
    • Mapping ideas from AlphaZero to what we learned during workshop

Dates

If you wish to enroll in this course please contact us on info@mlcollege.com.