An agent observes a state, picks an action, receives a reward, and repeats. That simple loop — run for millions of iterations — is how machines learn to play chess, control robots, and optimise ads.
A random agent navigates a 5×5 grid. ■ = agent, ■ = goal, ■ = obstacle.
Learning by doing: the agent discovers what works by trying actions and observing consequences.
Reinforcement learning (RL) is a framework for sequential decision-making. An agent interacts with an environment over time: at each step it observes the current state, selects an action, receives a scalar reward, and transitions to a new state. The agent’s goal is to maximise the cumulative reward over time — not just the next reward, but the long-run total.
Unlike supervised learning, there are no labelled examples. The agent learns purely from the consequences of its own actions, which may be delayed by many steps from the moment they were taken.
Every RL algorithm repeats the same four-step cycle:
Run this loop for millions of steps and the agent gradually finds behaviours that accumulate high reward — even if no single step’s reward told it what to do.
Supervised learning needs a dataset of (input, correct-output) pairs — someone had to label the training data. RL receives only a scalar reward, and that reward can be sparse (only at episode end: win/lose) and delayed (the action that caused the win happened 200 steps ago).
The credit assignment problem is the challenge of figuring out which actions in a long sequence were responsible for a good or bad outcome. Supervised learning sidesteps this entirely by providing a direct target at every step. RL must solve it from scratch.
RL algorithms split along two main axes:
At any moment the agent faces a dilemma: exploit what it already knows to earn reward now, or explore unfamiliar actions that might turn out to be better. A purely greedy agent gets stuck in local optima. A purely random agent never uses what it has learned.
Every RL algorithm must balance this tradeoff. The grid-world visualisation above makes it vivid: the random agent wanders aimlessly, reaching the goal by chance. An agent that has learned to exploit knowledge reaches the goal in a straight line.
RL has produced some of the most dramatic results in AI:
The reward hypothesis, attributed to Sutton, states: all goals and purposes of an intelligent agent can be described as the maximisation of the expected value of the cumulative sum of a scalar reward signal.
This is a working assumption, not a proven theorem. It is powerful because it reduces the problem of goal-specification to the problem of reward-function design — difficult, but tractable. Whether every goal can be described this way remains an open philosophical question.
Say the answer out loud before you reveal it — recalling it is what makes it stick, and rereading it is not.
What is the difference between a reward and a return in RL?
Why is the exploration vs exploitation tradeoff fundamental to RL?
What distinguishes RL from supervised learning?
What is the reward hypothesis?
RL = agent + environment + reward signal. The agent observes state s, picks action a, receives reward r, transitions to s'. Goal: maximise cumulative discounted return G_t = r_t + γr_{t+1} + ... Key challenge: credit assignment (which actions caused good outcomes?) and exploration-exploitation tradeoff.