Step by Step
A
Agent — the dog (or AI) doing the learning
The learner itself, which takes actions and gradually improves its behavior based on the feedback it receives.
Example: a dog being trained to sit, or an AI learning to play a game — both are "agents" in the RL sense.
E
Environment — the world the agent acts within
Everything the agent interacts with and receives feedback from — the physical world for a real dog, or a simulated game world for an AI.
Example: a living room is the environment for a dog being trained; a game board is the environment for a game-playing AI.
R
Reward — the treat or punishment driving learning
The feedback signal — a treat (positive reward) or a correction (negative reward/penalty) — that teaches the agent which actions to repeat and which to avoid, with the goal of maximizing total reward over time.
Example: a dog getting a treat for correctly sitting on command learns to repeat that behavior; a game-playing AI gaining points for a winning move learns to favor similar moves in the future.
Applied Walkthrough
1
AlphaGo needs to learn how to play the board game Go well enough to beat world champions.
2
The Agent (AlphaGo) takes actions (placing stones) within the Environment (the Go board and rules).
3
After each game, the Reward signal (winning or losing) tells the agent how well its overall strategy performed.
4
Over millions of games of trial, error, and reward feedback, AlphaGo's policy improved to the point of beating world champions — no human ever labeled individual moves as "correct"; the agent learned purely from reward signals.
Exam Application
Exams test whether you can identify the three RL components (agent, environment, reward) in a described scenario, and whether you understand RL requires no labels at all — just trial, error, and feedback, which distinguishes it clearly from supervised learning. Expect real-world examples like AlphaGo and robotic locomotion as illustrative test cases.
⚠ Common Trap
The most common trap is confusing reinforcement learning with supervised learning because both involve some form of feedback. The key difference: supervised learning provides an explicit correct label for every single example upfront, while reinforcement learning only provides a reward signal after actions are taken, with no example ever being explicitly labeled as "the right one" in advance.
✓ Quick Self-Check
1. What are the three core components of reinforcement learning in the reward dog story?
Agent, Environment, and Reward.
Tap to reveal / hide
2. What role does the "Agent" play?
It's the learner that takes actions and improves its behavior based on feedback — like the dog being trained.
Tap to reveal / hide
3. What role does the "Environment" play?
It's the world the agent acts within and receives feedback from — like the living room for a dog, or the game board for an AI.
Tap to reveal / hide
4. Does reinforcement learning require labeled training examples?
No — it requires no labels at all, just trial, error, and reward/penalty feedback.
Tap to reveal / hide
5. Name a real-world example where reinforcement learning famously succeeded.
AlphaGo beating world Go/chess champions, or robots learning to walk through trial and error.
Tap to reveal / hide