What is reinforcement-learning?

This repository is a collection of short, readable Python examples that teach reinforcement learning, a branch of machine learning where a program learns by trial and error, receiving rewards for good actions and penalties for bad ones. Each algorithm gets its own self-contained file, so you can open one script and see exactly how that technique works without wading through a larger codebase. The examples are organized into four levels of difficulty. The first level uses a simplified grid world, a small map where a virtual agent learns to navigate by trying strategies like policy iteration, value iteration, Q-learning, and a few others. Each name refers to a specific mathematical recipe for deciding what action to take next. The second level applies those ideas to CartPole, a classic test where a program tries to balance a pole on a moving cart. Here the examples introduce deeper techniques: Deep Q Network (DQN), Double DQN, Policy Gradient, Actor-Critic, and Asynchronous Advantage Actor-Critic (A3C). These combine the trial-and-error learning approach with neural networks, which let the program handle more complex situations. The third level moves to Atari video games like Breakout and Pong, showing how the same algorithms can learn to play from raw screen pixels. A fourth section covering OpenAI Gym environments is listed as work in progress. To run any example you need Python 3.5, TensorFlow, Keras, and a handful of standard scientific libraries. Installation is a single pip command. The project is maintained by a small team and welcomes pull requests and issue reports.