Reinforcement learning
Reinforcement learning is like training a pet by giving it a treat when it does something right and ignoring it or correcting it when it does something wrong, so it learns to repeat the good behaviors to get more rewards.
A machine learning paradigm where an agent learns to make decisions by interacting with an environment, receiving feedback in the form of rewards or penalties, and iteratively refining its strategy to maximize long-term cumulative gain.
A learning paradigm in which an agent takes actions in an environment and updates its policy to maximise cumulative reward; the basis of RLHF and many agentic systems.
evolution
- 1951 · historySNARC machine
Marvin Minsky and Dean Edmonds built the first neural network computer, which used reinforcement principles to simulate a rat in a maze.
- 1989 · historyQ-Learning
Chris Watkins introduced Q-learning, a model-free reinforcement learning algorithm that became a cornerstone of the field.
- 1992 · historyTD-Gammon
Gerald Tesauro developed a program that used temporal difference learning to reach near-expert level performance in backgammon.
- 2013 · historyDeep Q-Network (DQN)
DeepMind combined reinforcement learning with deep neural networks, enabling agents to learn directly from raw pixel inputs.
- 2017 · historyProximal Policy Optimization (PPO)
OpenAI introduced PPO, which became the industry standard for stable and efficient reinforcement learning training.