How to spot an RL-shaped problem
Before touring applications, learn the silhouette. A problem fits reinforcement learning when three things line up: decisions are made in a sequence over time (sequential decision-making), each decision affects what comes next, and there's a clear notion of better or worse outcomes you can score with a reward, giving the agent something to aim at — true goal-directed behaviour. If your problem has just one independent decision with a known right answer, you probably want plain supervised learning instead.
The applications of RL all share that shape. Let's walk through the big ones.
Games: the proving ground
Games were RL's first headline triumphs, and it's easy to see why: they're cheap to simulate, have crisp rewards (win/lose, score), and are clearly episodic. RL agents learned to play classic arcade games straight from the pixels, and AlphaGo beat the world's best Go players by combining learning with planning — a watershed moment that convinced many that RL could crack problems long thought to need human intuition.
An interactive search tree expanding the most promising branches of game moves.
Robots and control: acting in the physical world
Robotics is RL's natural home: a robot is literally an agent taking actions in an environment to reach a goal. RL has taught robots to walk, to grasp and manipulate objects, and to recover from a stumble. The same logic extends to control problems with no robot at all — steering datacentre cooling, managing power grids, scheduling traffic — anywhere a system makes a stream of decisions whose effects unfold over time, often as continuing tasks that never reset.
An interactive gridworld where a Q-learning agent moves toward a goal cell.
The physical world raises a hard wrinkle: real trial and error is slow, expensive, and sometimes unsafe. So much robot RL trains first in simulation and then transfers to hardware — a theme later tracks explore in depth.
Chatbots: RL inside the language models you use
Here's a surprise for newcomers: the helpful, polite tone of modern AI assistants is shaped by RL. After a large language model learns to predict text, it is refined with reinforcement learning from human feedback (RLHF): people rate which of two responses they prefer, those ratings become a reward signal, and the model is tuned to produce answers that earn more of it. The 'environment' is the conversation, the 'action' is the text generated, and the reward reflects human preference.
A pipeline where human preference comparisons train a reward model that fine-tunes the language model's policy.
When not to reach for RL
Honesty matters. RL is powerful but often the hardest tool in the box, and it isn't always the right one. Skip it when a simpler method will do, especially in these cases:
- If each decision is independent and you already have labelled right answers, use supervised learning — it's simpler and more reliable.
- If you have a perfect model of the world and just need the best plan, classical planning or optimisation may beat learning.
- If trial and error is too costly or unsafe and you can't simulate it, the exploration RL relies on may simply be off the table.
- If you can't define a meaningful reward, RL has nothing to aim at — fix the goal first.
Diagram of the supervised-learning loop: data to model to prediction to loss to update.
That's the whole 'What RL Is' track. You now know the core idea, the loop and its three signals, why long-term and delayed reward make learning hard, how RL stands apart from supervised learning and planning, and where it genuinely earns its keep. From here the domain dives into the mathematics — but every formula ahead is just a precise version of an intuition you now already hold.