Back to the index
42/ 55

HOW MODELS LEARN

Reinforcement learning.

RL

A way of learning through feedback in which an agent improves its choices by seeking higher cumulative reward.

In plain words

Reinforcement learning is a branch of machine learning in which an agent learns what actions to take by interacting with an environment and receiving feedback, often called rewards.

A closer look

At each step, an agent observes a situation, chooses an action, and receives a result: a new situation and a reward signal. The central challenge is credit assignment: working out which earlier choices helped or hurt a later outcome. A choice that has a small immediate cost can still be valuable if it improves the long-term result.

An agent also has to balance exploration and exploitation. Exploration tries actions that may reveal better strategies; exploitation uses what the system already believes works. The reward function matters enormously, because it defines what the learner is pushed to optimize. Poorly designed rewards can encourage shortcuts that score well without achieving the intended goal.

In practice

AN EXAMPLE

A game-playing agent receives a positive reward for winning and a negative reward for losing. After many games, it adjusts its strategy to make winning outcomes more likely, even though it was never given a list of correct moves for every position.

A useful distinction

Reinforcement learning does not mean a system learns exactly what its designers want. It learns to optimize the reward signal it receives, which can differ from the real-world goal. It is also distinct from reinforcement learning from human feedback (RLHF), which uses human preferences as part of the feedback process.

Sources & further reading

Sutton and Barto — Reinforcement Learning: An Introduction (opens in a new tab)