MMOOZZGG.com

Glossary

What is reinforcement learning?

In short

Reinforcement learning (RL) is a way of training in which a program tries actions and learns from a reward for good steps.

AI basics · reinforcement learning

MMOOZZGG.com editorsChecked: 1 min read

For a child

When you teach a puppy to “sit”, you do not explain anything in words. The puppy does different things: it jumps, barks and spins. As soon as it sits, its owner gives it a piece of cheese. After a few days the puppy sits right away, because it has learned what earns praise. You can teach a computer in a similar way: it tries different moves, gets points for good ones and slowly picks what brings more points.

In plain words

There are no ready right answers here. There is a program called an agent, an environment where it acts, and a reward: a number that says how well things went. The agent takes a step, sees the result, gets a plus or a minus and adjusts its strategy. The reward may not come at once: in chess, you know whether a move was good only at the end of the game. Computers learned this way to play video games, control robots and pick answers that people like.

For business

The method fits where a decision is made many times in a row and the result depends on a chain of steps: inventory control, dynamic pricing, courier routes, equipment tuning. There is also a use that affects every company with chat models: their answers reach a polite and useful tone through training on human ratings (RLHF). Companies rarely build this approach in-house. You need a process simulator or a long, safe history, or the model will learn on live customers and money. The main risk: the agent finds a loophole and grows the reward, not the profit. So agree on the reward formula and the limits of action before the start and test them in a pilot.

For a specialist

Reinforcement learning is formalized as a Markov decision process: states, actions, transition probabilities and a reward function. The agent looks for a policy that maximizes the expected discounted sum of rewards. Approaches include value-based methods such as Q-learning, policy gradient methods and actor-critic, which combines them. Algorithms are also split into model-free and model-based ones. The central problem is the balance of exploration and exploitation. The standard textbook is Sutton, Barto, “Reinforcement Learning: An Introduction”, second edition, MIT Press, 2018. Deep RL gained speed after the 2013 work by Mnih et al.: a convolutional network with a variant of Q-learning learned to play seven Atari 2600 games from raw pixels and beat a human expert in three. In language models, the method aligns models with human feedback, as in InstructGPT (Ouyang et al., 2022).

Example

A logistics company trains an agent on a simulator of its warehouse. The agent decides when and how much stock to reorder, and it gets a reward for fulfilled orders minus storage costs. After thousands of simulated months, the strategy moves to the real purchasing system under a manager’s control.

Related terms

Questions and answers

How is reinforcement learning different from supervised learning? In supervised learning, the model sees the right answer for each example. Here there is no right answer, only a score for the result, often delayed, and the model looks for the actions that lead to it on its own.
How is this method linked to chatbots? A ready language model is fine-tuned on human ratings: people compare answer options, and the model learns to give the ones that get high ratings. This method is called RLHF, and it makes answers more useful and safer.

Sources

Not sure whether your company needs this?

AI solutions map
Contact us