From Prediction to Decision: A Gentle Introduction to Reinforcement Learning

Last update: February 2026. All opinions are my own.

Overview

Reinforcement learning (RL) is the science of decision-making.

Unlike most machine learning, it is not about predicting labels. It is about learning how to act.

If supervised learning says "here is the right answer," RL says "try something and see what happens."

This post is a short, visual-first intro so you can read papers and code with confidence.

1) From prediction to decision

Most ML problems are about mapping inputs to outputs. RL is about choosing actions over time.

In RL, actions change what you see next. The data depends on the policy, and the policy keeps changing. That loop is the whole point.

A quick comparison

2) How RL differs from supervised and unsupervised learning

To understand RL clearly, compare it to the other two paradigms.

Supervised learning

You are given:

input -> correct output

The model's job is to minimize prediction error.

Examples:

• Spam detection
• Image classification
• House price prediction

You measure performance with accuracy, loss, MSE, and related metrics. The dataset is fixed.

Unsupervised learning

You are given:

input only

The model discovers structure:

• clusters
• patterns
• low-dimensional representations

There is no "correct answer." But the dataset is still static.

Reinforcement learning

You are given:

state -> choose action -> receive reward

There are no labels, no predefined dataset, and no immediate error signal.

Instead:

• The agent collects its own data.
• Its actions influence what it sees next.
• The goal is to maximize long-term reward, not minimize prediction error.

Think of how a baby learns to walk:

• Try a step.
• Fall.
• Adjust.
• Try again.

No labels. Just feedback.

Reinforcement learning is about decisions, not predictions.

LunarLander is not a prediction problem. It is a decision-making problem. The agent does not predict where the lander should be. It learns which engines to fire and when to land safely.

3) The RL loop

Every RL system is the same loop:

state -> action -> reward -> next state -> ...

The agent tries something, gets feedback, and updates its policy.

4) The core pieces (plain language)

You will see these terms everywhere:

• Agent: the decision-maker.
• Environment: the world the agent interacts with.
• State: what the agent observes.
• Action: what the agent does.
• Reward: the feedback signal.
• Policy: the rule that maps state to action.
• Episode: one full run from start to finish.

If the policy is the behavior, the value function is the prediction of how good a state or action is.

5) A tiny example

Imagine a robot in a maze:

• Each move is an action.
• The goal is to reach the exit.
• The reward is +1 at the exit, 0 elsewhere.

At first, the robot wanders randomly. Over time, it learns which paths lead to the exit and repeats them more often.

6) Why RL is hard in practice

RL is powerful but fragile:

• Rewards can be sparse or noisy.
• Exploration can be expensive or unsafe.
• Small reward tweaks can change behavior a lot.

7) Where to go next

If you want to go deeper, try these steps:

1. Build a small RL agent for CartPole.
2. Plot reward curves and watch for instability.
3. Read Sutton and Barto, Chapters 1-3.
4. Study PPO to see how modern agents learn reliably.

RL is a big field, but the core idea is simple: learn to make good decisions through experience.