Reinforcement Learning

A bandit decides one step at a time; reinforcement learning gives the agent a sequence of states and lets it learn by interacting for reward. The page starts with books, platforms, and lectures, then sketches Q-learning, moves to deep RL, and ends with RLHF.

The same notes are in Contextual Bandits and Multi Armed Bandits.

Introduction

The entry point is the textbook in both of its states. Reinforcement Learning: An Introduction second edition WIP is the second edition while it was still in progress, and the completed book is the finished version.

Beside the book, Vidhya has two Faizan Shaikh articles. Vidhya on Getting ready for AI based gaming agents – Overview of Open Source Reinforcement Learning Platforms explains reinforcement learning with examples and tours the major platforms such as Deepmind Lab and OpenAI Gym. Vidhya on Simple Beginner’s guide to Reinforcement Learning & its implementation defines reinforcement learning as a type of machine learning where an agent learns to maximize reward by interacting with an environment.

The reward idea also works outside games. ZipRecruiter on Classifying Job Titles With Noisy Labels Using REINFORCE — Fine-grained job title classification with noisy labels using the REINFORCE algorithm and multi-task learning.

The same notes are in Unbalanced labels.

The reason it is here: this article has a very nice trick in adding a reward component to the loss function in order to mitigate for unbalanced class label problem, instead of the usual balancing.

For a full course, David Silver — Home Page is his Teaching page, and the Google DeepMind recordings of his RL Course are the lectures in order:

1. 1: Introduction to Reinforcement Learning

2. 2: Markov Decision Process

3. 3: Planning by Dynamic Programming

4. 4: Model-Free Prediction

5. 5: Model Free Control

6. 6: Value Function Approximation

7. 7: Policy Gradient Methods

8. 8: Integrating Learning and Planning

9. 9: Exploration and Exploitation

10. 10: Classic Games

The figure is from those David Silver teaching materials.

David Silver teaching materials.

The last introductory source frames RL as sequential decisions in general. Sequential Decision Analytics and Modeling is the second edition from CASTLE, by Warren B. Powell, Professor Emeritus at Princeton University.

Q-LEARN

With the lectures behind you, Q-learning is the first algorithm to sketch. The problem is a Markov chain problem, (state, action, new state, reward). There is lots of exploration in the beginning, then exploitation, and the method returns optimal policy. For what Q itself is, the Udacity video What is Q is on youtube here.

Deep Learning

From Q-learning, the next step is RL combined with deep learning, deep Q-learning first. A review paper about RL in DL is Deep Reinforcement Learning: An Overview on arXiv. deep Q-learning is Ankit Choudhary's An Introduction To Deep Reinforcement Learning, which builds a deep Q-learning model in Python using keras and gym. The Pytorch tutorials then give runnable versions: DQN is the Reinforcement Learning (DQN) Tutorial, PPO is Reinforcement Learning (PPO) with TorchRL, and the Mario example trains a Mario-playing RL agent.

RLHF

After the RL introduction above, the same reward loop is used to tune language models from human feedback. The same notes are in Reinforcement Learning for LLM.

illustrated RLHF by Huggingface is Illustrating Reinforcement Learning from Human Feedback (RLHF), from a site that describes itself as on a journey to advance and democratize artificial intelligence through open source and open science.