Reinforcement Learning
Reinforcement learning is a branch of machine learning where a program, called an agent, learns to make good decisions by acting in an environment and receiving rewards or penalties for the outcomes, instead of being shown correct answers in advance.
itArtificial intelligence and machine learning | OpenSkills.info
Course pathWalk it in order
Look it upDip in anytime
Go furtherLeaves this page
Intro
Reinforcement Learning
Reinforcement learning trains a decision maker, called an agent, by letting it act in an environment and observing what happens. The agent takes an action, the environment returns a new situation and a reward, and the agent uses that experience to act better next time. Nobody hands the agent a labeled example of the right move.
That setup separates reinforcement learning from supervised learning. A supervised model learns from a fixed dataset of correct answers. An agent learns from its own trial and error, and its own choices change what data it sees next.
What it is for
Use reinforcement learning when the right behavior is a sequence of decisions, each one shaping the situations that follow, and when you can define a reward signal for good outcomes. Common use cases include:
Continue the course
This section is part of the paid course.
See pricing to subscribe, or log in if you already have access.
Where this skill leads
Relevant careers
See how this topic contributes to broader role-level skill maps.
Sources
- https://davidstarsilver.wordpress.com/teaching/
Supports
- Course structure: MDPs, planning by dynamic programming, model-free prediction, model-free control, value function approximation, policy gradient methods
- Epsilon-greedy exploration strategy (random action with probability epsilon, else best-known action)
- Multi-armed bandit as the single-state simplification of the exploration/exploitation tradeoff
- http://incompleteideas.net/book/the-book-2nd.html
Supports
- Canonical MDP formalism built from the multi-armed bandit case up to full state transitions
- Textbook is freely available, authored by Sutton and Barto, MIT Press second edition 2018
- https://spinningup.openai.com/en/latest/spinningup/rl_intro.html
Supports
- Definitions of state vs. observation, action spaces, deterministic vs. stochastic policy
- Return (finite-horizon undiscounted, infinite-horizon discounted with discount factor)
- Value function V, action-value function Q, optimal V*/Q*, advantage function A = Q - V
- Bellman equation informal statement (value of a state = expected reward + discounted value of next state)
- https://spinningup.openai.com/en/latest/spinningup/rl_intro2.html
Supports
- Model-based vs. model-free split
- Policy optimization family (on-policy: A2C/A3C, TRPO, PPO) vs. Q-learning family (off-policy: DQN, C51)
- On-policy vs. off-policy tradeoff (stability vs. sample efficiency)
- DDPG, TD3, SAC as off-policy actor-critic algorithms
- https://gymnasium.farama.org/introduction/basic_usage/
Supports
- env.reset() returns (observation, info)
- env.step(action) returns (observation, reward, terminated, truncated, info)
- One action-observation exchange is a timestep; episode ends when terminated or truncated is True
- https://deepmind.google/blog/alphago-zero-starting-from-scratch/
Supports
- AlphaGo Zero trained solely by self-play, combining a neural network with a search algorithm, no human game data
- Network updated to predict moves and the eventual winner
- AlphaGo Zero defeated the champion-defeating AlphaGo version 100 games to 0 after 3 days of training
- Authored by David Silver and Demis Hassabis, published by Google DeepMind, October 18, 2017
- https://github.com/awesomelistsio/awesome-reinforcement-learning
Supports
- Stable-Baselines3, Ray RLlib, PettingZoo, Unity ML-Agents, TF-Agents as curated RL frameworks/tools
- https://api.github.com/repos/DLR-RM/stable-baselines3
Supports
- Recency check confirming Stable-Baselines3, Ray RLlib, PettingZoo, Unity ML-Agents, and TF-Agents all had commits pushed within 2026 and are not archived