---
title: Reinforcement Learning
type: vocabulary
url: "https://www.envisioning.com/vocab/reinforcement-learning"
summary: A machine-learning paradigm in which an agent learns to take actions in an environment by maximizing cumulative reward through trial and error, rather than from labeled examples.
year: 1954
generality: 0.90
---

# Reinforcement Learning

A machine-learning paradigm in which an agent learns to take actions in an environment by maximizing cumulative reward through trial and error, rather than from labeled examples.
Reinforcement learning (RL) is a machine-learning paradigm in which an agent learns a policy, a mapping from states to actions, by interacting with an environment and receiving a scalar reward signal. This differs from supervised learning, which trains on labeled input-output pairs. The agent tries to choose actions that maximize cumulative reward over time. Doing so requires balancing exploration of unfamiliar actions against exploitation of actions already known to work well. Early computational studies of trial-and-error learning include Claude Shannon's 1952 maze-solving mouse, Theseus, and Marvin Minsky's 1954 doctoral work on an analog learning machine he called SNARC.

Minsky's 1961 paper "Steps Toward Artificial Intelligence" named the credit-assignment problem: determining which past decision deserves credit for a later outcome. It remains central to the field. Research on trial-and-error learning was largely dormant through the 1960s and 1970s before reviving in the early 1980s, driven by temporal-difference methods and, later, Q-learning (Watkins, 1989). Richard Sutton and Andrew Barto's textbook "Reinforcement Learning: An Introduction" (MIT Press, 1998) consolidated the field's mathematical framework around Markov decision processes, value functions, and temporal-difference learning. This vocabulary remains standard.

Deep reinforcement learning combines this framework with neural networks as function approximators, letting an agent handle high-dimensional observations such as pixels or text. Notable results include DeepMind's Deep Q-Network playing Atari games from raw pixels (2015) and AlphaGo defeating a professional Go player (2016). Beyond games and robotics, reinforcement learning underlies reinforcement learning from human feedback (RLHF) and reinforcement learning with verifiable rewards (RLVR), both used to align large language models. It also underlies policy-gradient methods for tasks like neural architecture search, where a controller network is rewarded for designing better neural network architectures.

---
Source: Envisioning — Technology Research Institute (https://www.envisioning.com/vocab/reinforcement-learning)