dennybritz/reinforcement-learning

This repository is a learning resource for reinforcement learning, a branch of artificial intelligence where a software agent learns to make decisions by trial and error, receiving rewards for good actions and penalties for bad ones. Think of it like training a dog with treats, but applied to algorithms. The code is designed to accompany two specific learning materials: the textbook "Reinforcement Learning: An Introduction" (2nd edition) by Sutton and Barto, and David Silver's university lecture course on reinforcement learning. Each folder in the repo corresponds to a chapter or topic from those materials, and contains exercises, worked solutions, a summary of the key concepts, and links to further reading. The implemented algorithms cover a progression from foundational to more advanced techniques: dynamic programming (planning when you have a complete model of the environment), Monte Carlo methods (learning from complete episodes of experience), temporal difference learning (learning step by step without waiting for an episode to end), Q-Learning (a widely studied off-policy method), and Deep Q-Learning (combining Q-Learning with neural networks to handle complex problems like Atari games). Policy gradient methods and an actor-critic algorithm are also included. Everything is written in Python 3 using Jupyter Notebooks, interactive documents that mix code, explanations, and output, and uses OpenAI Gym for training environments and TensorFlow for the neural network-based algorithms. You would use this repo if you are studying reinforcement learning and want hands-on code alongside the theory.