Learn AI /How Machines Learn /Reinforcement learning

Reinforcement learningRL

Intermediate
How Machines Learn
Technical readers
Policy officers

Also called: RL.

Detail

In five seconds

What it is

A training method that aims to improve behaviour through trial, feedback and cumulative reward across a sequence of actions.

In plain words

The system learns by trying things, seeing what leads to better results over time, and adjusting what it does next.

An everyday example

A new staff member helps sort incoming messages into an urgent tray or a later tray. At first, they guess and sometimes place a message in the wrong tray. Later in the day, they learn which choices helped the team respond quickly and which choices caused delays, and they adjust how they sort the next messages. After many days, they become better at recognising patterns that usually lead to faster handling overall, even when one message is hard to judge by itself.

Where the story stops working

A human staff member can use background knowledge, office norms and explicit explanations from colleagues. A reinforcement learning system does not rely on that kind of ordinary human reasoning unless those signals are built into the training setup. Also, real reinforcement learning often learns from thousands or millions of repeated interactions in a formal environment, not from a small number of daily examples.

Remember thisReinforcement learning improves a policy through rewards received over sequences of actions. The best action now may be the one that preserves better options, and higher cumulative reward, later.

Understand it

Why it matters

For non-specialists, reinforcement learning matters because it is used in situations where choices affect later options, costs or risks. That makes it relevant to resource allocation, scheduling, recommendations, control systems and some forms of AI-assisted optimisation. If you are assessing a system, you need to ask different questions than you would for ordinary prediction tools: What reward is it optimising? Who defined that reward? Does short-term success hide long-term harm? What happens while it explores and makes mistakes? Could it learn behaviour that scores well but conflicts with safety, fairness or policy goals? Understanding this helps with procurement, oversight and realistic expectations.

How reinforcement learning improves behaviour over repeated rounds

Show that reinforcement learning uses a loop of state, action, reward and update to build a better policy over time.

01
1. Observe state
The system receives information about the current situation, such as position, recent events or resource levels.
Draw a rounded box labelled 'Current state' with 2 to 4 small example icons or text chips inside, such as queue length, location or time.
02
2. Choose action
Using its current policy, the system selects one action from the available options.
Draw an arrow to a second box labelled 'Action' with 3 small buttons inside to suggest choices; highlight one selected button.
03
3. Environment responds
After the action, the situation changes. The system may reach a better, worse or neutral outcome.
Draw a third box labelled 'New state' with changed icons compared with step 1, plus a small transition arrow from action to new state.
04
4. Receive reward
A reward signal gives feedback on how useful the recent outcome was for the overall goal.
Draw a small score panel beside the new state showing examples like +2, 0 or -1 in clear text, with colour not being the only cue.
05
5. Update policy
The system adjusts its action-selection rule so better long-term actions become more likely in similar states.
Draw a final box labelled 'Policy update' with a slider or weight bars shifting slightly, then a curved arrow returning to step 1 to show repetition.

A left-to-right loop shown as five connected steps. Read from the current situation to action, then to feedback, then to policy update, then back to a new situation. A small side note shows that rewards may reflect long-term results, not just the immediate outcome.

On hover or tap, each step reveals one sentence explaining what information enters and what changes before the next step.

Describe this without the diagram

The diagram presents reinforcement learning as a repeating five-step process. First, the system observes the current state of the environment. Second, it chooses an action using its current policy. Third, the environment changes and produces a new state. Fourth, a reward signal indicates whether the outcome was better, worse or neutral for the goal. Fifth, the system updates its policy so actions linked to higher long-term reward become more likely in similar states. A return arrow shows that this cycle repeats many times.

With reduced motion: Show the same five boxes in a static row with a visible return arrow from policy update back to current state. Use labels and text only, with no animation.

Reinforcement learning trains a system by letting it try actions in a situation and then giving it feedback as a reward, a score that signals how useful the outcome was. The system is often called an agent, and the situation it acts in is called the environment. The goal is not just one good move, but a strategy that does well over time.

What goes in is usually a description of the current state, meaning the information available about the situation right now, plus a reward signal after actions are taken. Unlike supervised learning, there usually is not a correct answer attached to each example in advance. Instead, the system must discover which actions lead to better long-term results.

What happens inside is repeated trial and adjustment. The system takes an action, the environment changes, and a reward arrives. Over many rounds, it updates a policy or value estimates used to choose actions. The aim is to favour actions expected to lead to more cumulative reward in similar situations.

What comes out is a policy or control strategy, not just a label or a score for one item. That is why reinforcement learning is commonly used for problems with sequences, timing and trade-offs, such as routing, game playing or managing resources. It is different from unsupervised learning too, which looks for patterns without reward signals tied to actions.

How “Learning how to sort urgent messages” maps onto the real thing

In the story	In the AI system
staff member sorting messages	agent choosing actions
incoming message and current queue	state of the environment
urgent tray or later tray	available actions
learning from delays or quick responses	reward feedback used to update behaviour

A worked example

A warehouse trains a system to route one parcel at a time through either a fast conveyor or a regular conveyor. Each routing decision changes the queues faced by later parcels.

Starting from: State: both queue lengths, parcel size and remaining loading-dock capacity. Actions: fast conveyor or regular conveyor. Reward: +3 for on-time arrival, -2 if the choice pushes a conveyor above its operating queue, and -3 if loading is blocked.

The state is: eight parcels waiting, a small parcel ready, the fast conveyor moderately busy and the loading dock nearly full.
The system chooses the fast conveyor. The parcel arrives on time, earning +3, but the choice pushes that conveyor above its operating queue, adding -2. The immediate reward is therefore +1.
The next state now includes a crowded fast conveyor. If later parcels are delayed or loading is blocked, those lower future rewards reduce the cumulative return associated with the earlier choice.
Across many simulated sequences, training updates the policy or its estimates of future reward for each state and action.
The policy gradually learns that sending every small parcel to the fast conveyor can produce a lower cumulative reward than preserving capacity for later parcels.

Result: The trained policy does not merely select the conveyor that looks fastest now. It favours actions expected to produce the best cumulative reward across the sequence, including consequences that appear several routing decisions later.

In a United Nations workplace

An ombudsman case officer, sitting in a UN Country Team working group, receives a side-event concept note for clearance by close of business. It proposes demonstrating a simulator that learns which sequence of prompts to offer managers in fictional workplace-conflict exercises. The simulated agent observes how each conversation changes, selects the next prompt and receives rewards for outcomes such as keeping options open, making an appropriate referral and avoiding escalation.

There is no single labelled correct prompt at each point. An early choice changes the later options, so training uses cumulative reward across the simulated conversation to update the policy. The case officer asks the drafters to state clearly that the system learns in simulation, not from live ombudsman work.

Before you try thisConfidential case records, identities and free-text accounts must not be entered, replayed or used as feedback in any system not expressly approved for that material. Doing so could expose individuals and undermine confidentiality. Even with synthetic data, reward design needs review: rewarding rapid closure, for example, could teach coercive or superficial behaviour rather than sound handling.

Where you may hear it

At a vendor demonstration to an ICT governance board
During first reading of a concept note for an adaptive system
In a procurement unit meeting on an optimisation prototype
At an evaluation office review of a model's reward design

Avoid confusion

Often confused with

Common misconceptions

Common misconception: It just means rewarding a model whenever it gives the right answer.
Correction: That describes only a small part of the idea and can sound like supervised learning with scores. In reinforcement learning, feedback is often delayed, actions affect future states, and the main problem is choosing a sequence of actions that leads to better long-term reward.
Common misconception: A reward function automatically captures what people really want.
Correction: A reward function is only a designed signal. If it is too narrow, the system may learn behaviour that scores well on the reward while missing the real goal, creating side effects or exploiting loopholes in the setup.
Common misconception: It learns only by random guessing.
Correction: Exploration often includes trying uncertain actions, but practical methods usually balance exploration with exploitation, meaning using actions that already seem promising. The learning process is guided by the reward structure and update rules, not pure randomness.
Common misconception: Reinforcement learning from human feedback is the same thing as all reinforcement learning.
Correction: Reinforcement learning from human feedback is a specific approach where human preferences help shape the reward signal. Reinforcement learning more broadly covers many settings, including simulated environments and control problems without direct human ratings.

Quick check

Which situation best fits reinforcement learning rather than ordinary supervised learning?

Go deeper

The simple version

The system learns by trying things, seeing what leads to better results over time, and adjusting what it does next.

The technical version

In reinforcement learning, an agent interacts with an environment over time steps. At each step it observes a state, selects an action according to a policy, receives a reward and transitions to a new state. Training updates the policy, or a value function estimating expected future reward, with the aim of improving expected cumulative reward rather than matching pre-labelled answers. This differs from supervised learning because there is usually no fixed target output for each input, and feedback may be sparse or delayed. It also differs from standard prediction tasks because the agent's actions change the future data it will encounter. Many formulations use the language of Markov decision processes, but not all practical systems fit that framework perfectly.

Reviewed external resources

EncyclopaediaReinforcement learning (opens in a new tab)Wikipedia
This link gives a broad reference-style overview of reinforcement learning, including key terms and history, and suits readers who want a general starting point.
Intermediatelink checked 2026-07-31
CourseMachine Learning Crash Course (opens in a new tab)Google
This link offers a structured introductory course on machine learning concepts and suits readers who prefer guided lessons over reference pages.
Intermediatelink checked 2026-07-31
VideoStatQuest (opens in a new tab)StatQuest with Josh Starmer
This link provides an explainer-style video treatment from StatQuest and suits readers who learn best from clear verbal walkthroughs.
Starterlink checked 2026-07-31
VideoVideo explanations: Reinforcement learning (opens in a new tab)YouTube search
This link is a YouTube search results page, not one single explanation, so readers can compare different video styles and difficulty levels.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: Reinforcement learning (opens in a new tab)Reddit search
This link is a Reddit search results page, not one single discussion, so readers can browse community questions, informal explanations and debates about the term.
search results, not one pagelink checked 2026-07-31

Where this came from

Drafted with a language model through an offline pipeline, checked against this site's schema, and every external link above fetched to see whether it answers. Treat it as a starting point and check anything that matters against the primary sources. How this site is built and what that means.

The one thing to remember

Reinforcement learning improves a policy through rewards received over sequences of actions. The best action now may be the one that preserves better options, and higher cumulative reward, later.