A machine-learning technique where agents learn to take actions to maximise cumulative rewards through trial and error.
Also known as: RL
In plain English
Reinforcement learning is like training a dog with treats: the system tries actions and gets rewards or penalties and gradually learns what works. Over time it prefers actions that earn higher rewards without anyone giving exact step-by-step instructions.
In practice
In products, reinforcement learning shows up when systems optimise behaviour from interaction—examples include personalised recommendations, game AI and robotics, or chatbots tuned with user feedback. It matters because it can improve long-term outcomes but requires clear reward signals, many interactions or simulations and often higher compute and engineering effort than supervised learning.
Under the hood
Formally, reinforcement learning solves a Markov decision process by learning a policy π(a|s) to maximise expected cumulative reward, using value-based (e.g. Q-learning), policy gradient or actor‑critic methods, often with deep neural networks (deep RL). Key concerns are credit assignment, exploration versus exploitation, sample efficiency, stability of off‑policy learning and specification of the reward signal; RLHF is a common variant that uses human preference models as the reward.
Example
We should try reinforcement learning to optimise the recommendation engine for long-term user engagement.