How Computers Learn by Trying, Explained From Zero

The big idea

Reinforcement learning is how computers learn by trying. A computer system tries an action, sees what happens, and adjusts based on whether the result was good or bad. Given enough tries, it gets better. This is how some of the most powerful AI systems in the world — from chess-playing programs to the technology behind helpful chatbots — were built.

1 · You already understand reinforcement learning

Think about how you learned to do anything difficult — shoot a basketball, type on a keyboard, parallel park a car, cook a dish that actually tastes good. Nobody taught you by handing you a complete list of instructions for every possible situation you might encounter. You learned by doing it, noticing what happened, and adjusting next time.

When you shoot a basketball and it bounces off the rim to the left, you do not need anyone to tell you "you pushed slightly right." Your brain figures it out from the outcome and adjusts the next attempt. When the dish you cook comes out too salty, you use less salt next time. The outcome of each attempt teaches you something.

This basic mechanism — try, observe the result, adjust, repeat — is what computer scientists call reinforcement learning. The word "reinforcement" comes from psychology. When an animal does something and gets a reward, that behaviour gets reinforced — strengthened, made more likely to happen again. When it gets a punishment or simply nothing, the behaviour gets weakened. Over time, the animal learns which behaviours produce rewards.

AI researchers realised in the second half of the twentieth century that the same mechanism could be applied to computer programs. A computer program trying to play a game, control a robot, or recommend content could be set up to try things, observe results, and adjust based on whether the results were good or bad. It would not need to be told in advance exactly what to do in every situation. It could figure that out through experience.

2 · The three-part loop every RL system uses

Every reinforcement learning system, no matter how simple or complex, runs the same basic loop. It has three moving parts, and they go around and around:

Figure 1. The reinforcement learning loop. The agent takes an action, sees what happens, receives a reward signal, and uses that to decide what to do next. Over many repetitions, it learns which actions produce good outcomes.

The agent is the AI system doing the learning. Think of it as the student. It observes the current situation and decides what action to take.

The environment is everything the agent interacts with. In a chess game, the environment is the board and the rules. In a recommendation system, the environment is the user and everything the user can do. In a factory robot, the environment is the physical world around the robot arm.

The reward signal is the feedback the agent gets after taking an action. It is usually a simple number: a positive number if the action was good, a negative number if it was bad, and zero if it was neutral. The agent does not need a lengthy explanation of why the action was good or bad. The number is enough for it to learn over time.

The loop goes: observe the situation → take an action → see what happened → receive a reward number → use that to update understanding → observe the new situation → and repeat. Over thousands or millions of repetitions, the agent builds up an understanding of which actions tend to lead to good outcomes in which situations.

An agent takes an action, the environment changes, a reward returns, and the agent adjusts before trying again.

3 · The reward signal: the most important piece

If you had to pick the single most important concept in reinforcement learning, it would be the reward signal. It determines everything the agent learns.

Here is the critical point: the agent learns to do whatever produces the most reward, even if that is not what you intended. If you design the reward signal carelessly, the agent will figure out how to get high reward numbers in ways you never expected — and those ways might not be what you actually wanted.

Imagine you are teaching a dog to fetch a ball. You decide to reward the dog every time it comes back to you quickly after you throw the ball. Sounds reasonable. But the dog might learn that the fastest way to come back to you is to not run after the ball at all. It stays near you, gets the reward, and never fetches anything. The reward was "come back quickly," and the dog found the fastest way to do that. You wanted "fetch the ball and come back quickly," but that is not exactly what you told it to do.

This happens constantly in AI systems, and it is not the system's fault. The system is doing exactly what it was designed to do: maximise the reward signal. If the reward signal does not perfectly capture what you actually want, the system will optimise for the signal rather than for what you meant. This problem is important enough that later guides in this series are dedicated entirely to it. For now, just remember: the reward signal is the instruction that the AI actually follows, not the description of the goal you had in mind.

A careless reward teaches the agent to chase the number instead of the goal you meant.

4 · How this differs from other kinds of AI learning

There are three main ways AI systems learn, and they are quite different from each other:

Supervised learning works from examples. You show the system thousands of labelled examples — photographs labelled "cat" or "dog," emails labelled "spam" or "not spam" — and it learns to recognise the pattern. It needs someone to provide the right labels in advance. It cannot learn things that no human knows to label.

Unsupervised learning finds patterns in data without any labels. You give it a pile of data and it groups similar things together or finds structure that was not obvious before. It does not need labels, but it also does not have a specific goal to optimise for.

Reinforcement learning is different from both. The system does not need labelled examples. It does not need to be told the right answer. It tries things and learns from the results. This is what makes it powerful: it can learn strategies and solutions that no human figured out in advance, because it discovers them through its own experience of what works.

The tradeoff is that it needs to try things many, many times. A human child learning to walk takes a few months. An RL system learning to control a simulated robot might need millions of simulated attempts. Speed is not its strength; discovery is.

Reinforcement learning differs from supervised and unsupervised learning by learning from outcomes, not labels.

5 · Where reinforcement learning shows up in the real world

Reinforcement learning is not a theoretical curiosity. It is running in systems you interact with every day:

Game-playing AI. The AI systems that beat world champions at chess, Go, and video games were trained using reinforcement learning. They played millions of games against themselves, learned from wins and losses, and eventually discovered strategies that human players had never considered.
AI assistants and chatbots. The technology that makes a chatbot give helpful, appropriate answers — rather than just statistically plausible ones — involves a form of reinforcement learning where human feedback trains the system to behave better. The next guide in this series explains that in detail.
Recommendation systems. The systems that decide which videos, articles, posts, or products to show you are constantly learning from your behaviour. Your clicks, your time spent watching, your purchases — all of these are reward signals that the system uses to adjust what it shows you next.
Industrial robots. Manufacturing robots that need to perform precise, complex movements can be trained using reinforcement learning to learn those movements more effectively than human programmers could specify them manually.
Traffic and logistics. Systems that optimise traffic light timing, delivery routes, or power grid distribution use reinforcement learning to adapt to changing conditions in real time.

6 · Why understanding this matters

Reinforcement learning is not just something for AI researchers to understand. The systems it produces are shaping what information you see, how products are priced, what content is recommended to children, and increasingly, what decisions are made in areas like healthcare, finance, and hiring.

Understanding the basic mechanism — that these systems are optimising a numerical reward signal, not following a human-written set of rules — helps you ask better questions about them. When a recommendation system pushes content that seems extreme or divisive, understanding RL tells you why: extreme content may produce stronger engagement signals (clicks, time spent), and the system has learned to maximise those signals. When a hiring tool produces biased results, understanding RL helps you ask: what was in the reward signal that caused this? When a chatbot gives a confident but wrong answer, understanding RL helps you see why: the system learned what gets good feedback from humans, and confident answers often do.

None of this requires technical expertise. It requires understanding the basic loop and the fundamental role of the reward signal. That is what this guide has given you.

7 · Something to think about

A real-world example to think through

Think about a social media platform's recommendation algorithm. What is the "reward signal" it is optimising for? (Hint: it is something they can measure, like clicks or time spent.) Now think about whether maximising that signal might produce results the platform didn't intend. What would happen if users got more engagement from upsetting content than from informative content? What would the system learn to show?

There is no trick answer here. This is exactly the kind of thinking that people working in AI policy, product ethics, and AI safety do every day.

8 · Terms used in this guide

Reinforcement learning (RL): A way for AI systems to learn by trying things and receiving feedback about whether each attempt was good or bad. The system learns over many attempts to do more of what gets rewarded and less of what does not.
Agent: The AI system that is doing the learning and making decisions. In a game-playing AI, the agent is the program playing the game. In a recommendation system, it is the algorithm deciding what to show you.
Environment: Everything the agent interacts with. The game board in a chess game. The user and website in a recommendation system. The physical world around a robot arm. The environment receives the agent's actions and produces results.
Reward signal: A number given to the agent after each action, indicating how good or bad that action was. Positive numbers mean good outcomes. Negative numbers or zero mean bad or neutral outcomes. The agent learns to take actions that produce higher reward numbers over time.
Policy: The strategy the agent has learned. For any given situation, the policy tells the agent what action to take. Learning a good policy is the whole goal of reinforcement learning training.
Supervised learning: A different AI learning method where the system learns from labelled examples. A human provides the correct answers, and the system learns to match them. Useful but requires a human to label everything in advance.