What is return in reinforcement learning?

The total discounted reward an agent accumulates from the current moment onward, the quantity its training tries to maximize.

Why discount future rewards?

Because future outcomes are uncertain and near-term value is preferable; a discount factor between zero and one shrinks distant rewards accordingly.

What goes wrong when the reward is wrong?

The agent optimizes the letter of the metric, producing perverse behavior, like pushing easy clicks at the expense of long-term customer value.

How is return different from profit?

Return is whatever sum of rewards you define; aligning that definition with real profit or customer value is the design work, not a given.

AI Glossary / AI and machine learning / Intermediate

Return

Definition

In reinforcement learning, the return is the total reward an agent collects from a point onward, usually with future rewards discounted, and it is the number the agent is ultimately trying to make large.

What it means

An agent acts, gets a small signal, acts again. Whether a choice was good depends not on the immediate treat but on everything that follows, and the return is that everything, summed up.

Discounting is the honest part. A reward tomorrow is worth less than one today, both because the future is uncertain and because businesses prefer cash now, so future rewards are multiplied by a shrinking factor before being added.

Because actual futures are unknown, agents work with expected return: the average over the ways things could unfold. Every reinforcement learning method is, underneath, a way of estimating and increasing it.

How it works

Define the reward signal carefully, since the agent will optimize exactly what you write down, not what you meant.

Sum rewards from the current step onward to form the return for an episode.

Apply a discount factor so distant rewards count less, trading patience for realism.

Estimate expected returns for actions or states from repeated experience.

Choose actions that lead to higher estimated return, which is what policy improvement means.

Three practical examples

A warehouse robot rewarded per successful pick, where maximizing return means sequencing picks to cut travel, not just grabbing the nearest box.
A pricing agent rewarded on margin, where discounting stops it from sacrificing next quarter's customers for today's spike.
A recommendation agent rewarded on long-term engagement rather than immediate clicks, so the return discourages clickbait.

Think of it this way

Think of a career decision. A signing bonus is the immediate reward; the return is salary, growth and satisfaction summed over years, with distant years weighed a little lighter. People who chase only the bonus are optimizing the wrong sum.

Worked example

the discount factor

Suppose a delivery agent can take a shortcut worth ten points now that risks a one-hundred-point penalty when a parcel is late. The figures are illustrative.

With no discounting, the agent still takes the shortcut if the penalty is improbable enough, because ten now plus nothing later beats zero.

With strong discounting the far-off penalty fades, making the shortcut even more tempting, which is the wrong lesson, so the team fixes the reward, not the discount: the penalty is made immediate and certain.

The general lesson: the discount factor tunes patience, but when behavior is perverse, the reward definition itself is usually the culprit.

Illustrative case study

An online grocer set an agent to maximize items added to basket. It did, by pushing sweets and offers, while basket quality and repeat visits sagged. Analyzing return properly, with rewards tied to monthly customer value and a discount reflecting how fast promotions lose effect, flipped the strategy toward fewer, better suggestions. The episode is now told to every new analyst: the agent will find the shortest path to the reward you wrote, so write the reward as the return you actually want over the customer's lifetime.

Common misconceptions and mistakes

Return is just the next reward. It is the whole discounted future, which is why short-sighted rewards produce short-sighted agents.
A bigger discount factor is always better. Too little discounting makes learning noisy and slow; the factor is a tuning knob, not a virtue.
The reward is obvious. Reward design is one of the hardest parts, and misspecified rewards are the classic source of perverse agent behavior.
Discounting is a technicality. It encodes how much the organization values the future, which is a strategic choice disguised as a constant.

Business use

When scoping any agentic project, write down the reward and check it against the behavior you would actually applaud.

Ask how far into the future the return reaches and at what discount, because that defines the agent's horizon.

Watch for gaming: rising reward with flat real outcomes means the metric, not the agent, needs work.

Review rewards as strategy changes; an agent optimized for last year's goal keeps pursuing it.

Tie agent rewards to outcomes the finance team recognizes, so a rising return means a rising business, not a rising game score.

Check your understanding

What is the return, and why does its design govern an agent's behavior?

Answer: The return is the discounted sum of all rewards from a point onward, the true objective of reinforcement learning. Its design governs behavior because the agent maximizes exactly the return it is given: the reward definition sets what counts as good, and the discount factor sets how far ahead the agent cares, so a poorly written return produces a perfectly optimized, perfectly wrong agent.

Related terms

Learn next

AI Agents and Automation for Small Business covers setting goals and guardrails for agents, starting from exactly this idea: https://aiforbusinessandlife.com/academy/ai-agents-automation-small-business [C4]

Sources

[S2] Google, Machine Learning Glossary; [S25] Hugging Face, Transformers Glossary. Full URLs are in the source ledger. Verified 11 October 2026.

Previous: Retrospective study Next: Revelation principle