WHAT WE ARE DOING

WHY WE ARE DOING IT

We will tune an AI bot's parameters to give it a distinct personality, observe how it behaves in a dungeon maze, and document your findings in a field guide entry.

To understand reinforcement learning by watching it happen live in a game, then connecting it to real AI in the world.

؜2nd

3rd

4th

Play First, Think Later

Learn about Reinforced Learning

Design Your Bot

Build Your Field Guide

ACTIVITY STEPS

1st

5th

Build Your Own Game

What happens when it makes a wrong move?

Does the bot know where to go straight away?

Does its behavior seem to change over time?

QUESTIONS TO CONSIDER

1st

Go to

When the game loads, select the Reinforcement Learning mini-game from the main menu NOT the Supervised Learning one.

X min

Play First, Think Later

Play through it once without overthinking. Just observe.

Reinforcement learning is a type of machine learning where an AI learns by trying things, seeing what happens, and adjusting its behavior based on rewards and penalties.

The AI figures it out by exploring, making mistakes, and gradually doing better. It learns the same way you might learn a new video game: by playing, failing, and slowly figuring out what works.

REINFORCEMENT LEARNING

2nd

KEY INGREDIENTS

Episode

Policy

Agent

Environment

State

Action

Reward

Penalty

The AI that acts: Our ArtBot

The world it acts in: The maze

The bot's position

What the bot can do: ex. move up

A positive signal for good moves

A negative signal for bad moves

One full attempt through the maze

The strategy the bot develops

HOW DOES AN EPISODE WORK?

Update Knowledge

Repeat

؜Observe State

Choose Action

Get Feedback

The bot looks at where it is in the maze right now.

It picks a direction to move. Sometimes randomly, sometimes based on what it's learned.

The environment gives it a reward (good move) or a penalty (bad move).

The bot adjusts its understanding of which moves work in which situations.

The loop starts again from the new state. Over many episodes, the bot gets better.

EXPLORATION vs. EXPLOITATION

EXPLORE

EXPLOIT

Good RL balances both: explore enough to discover better paths, exploit enough to actually get good at the task.

OR

Try something new:

a direction it hasn't taken from this spot before.

It might find a better path, or it might find a wall.

Stick with what worked before:

repeat the action that got a good reward last time.

Safe, but it might miss a better route.

EXPLORATION RATE vs. LEARNING RATE

Exploration Rate

Learning Rate

Every RL bot has two main settings that shape how it learns:

AND

How often the bot tries a random action instead of the best one it knows.

High → always wandering, trying new things

Low → sticks rigidly to what worked before

How strongly each new experience updates the bot's beliefs.

High → overreacts to every move, unstable

Low → cautious, very slow to update, very steady

3rd

Go to

When the game loads, select the Reinforcement Learning mini-game from the main menu NOT the Supervised Learning one.

Design Your Bot

1. ؜Choose of one the 5 profiles to give your bot.

1. Set the parameters to match your creature's profile.

1. Run 5-10 episodes. Take notes on what you observe.

1. Compare: did the bot behave the way you predicted? What surprised you?

BOT PERSONALITIES

Cautious

Reckless

Explorer

Efficient

Random

؜We will be choosing one of these personalities to try to give to our bot

Plays it safe.

Rarely strays from the first path that worked.

Tries everything at full speed.

Unstable and impulsive.

Always wandering. Curious but never commits.

Steady improvement. Finds the best path and locks it in.

Pure chaos.

No learning, no memory, no strategy.

ASK YOURSELF

Cautious

Reckless

Explorer

Efficient

Random

how would each of those personalities affect the bot's learning and exploration slider?

Plays it safe.

Rarely strays from the first path that worked.

Tries everything at full speed.

Unstable and impulsive.

Always wandering. Curious but never commits.

Steady improvement. Finds the best path and locks it in.

Pure chaos.

No learning, no memory, no strategy.

For your bot, ask:

would exploration be high or low here?

what about learning?

would it care about a reward or a punishment?

؜Things ArtBot wants to find.

When ArtBot collects a reward, it receives positive feedback and learns that the path it took was a good choice.

1. Encounter object
2. Collect rewards to earn points.
3. Positive reward = "Do this again!"

The more rewards ArtBot finds, the better it learns where to go.

Game Rewards +

؜Traps

Things ArtBot should avoid.

If ArtBot steps on a trap, it receives negative feedback and learns that this path is not a good choice.

1. Encounter object
2. Lose points.
3. Negative reward = "Avoid this path."
4. After enough experience, ArtBot will try to avoid traps.

Game Traps -

؜The final goal.

1. Once ArtBot has collected rewards and avoided traps, it should find the exit as efficiently as possible.

Game Settings: Exit

1. Duration: Long: More time to learn and improve. Short: Less time, so it may not find the best path.
2. Exploration: High: Tries lots of new paths. Low: Sticks to routes that already worked.
3. Experience: High: Remembers where rewards were found. Low: Often forgets and starts fresh.
4. Learning Rate: High: Learns from rewards right away. Low: Needs to find the same reward several times before it remembers.

Game Settings

BOT EXAMPLES

Here are two Examples of different settings for the game's level one

Based on what we learned, how would you describe each of those AI bots?

Exploration: High

Learning Rate: Low

Exploration: Low

Learning Rate: High

4th

Document your creature below. This is YOUR scientific record .Write in your own words.

Build Your Field Guide

NAME YOUR BOT

؜BEFORE: Your Prediction

EVIDENCE

PARAMETER SETTINGS

AFTER: What You Observed

VERDICT

Give your specific bot a unique name + a one-line personality description.

What did you expect to happen? Describe episode 1 vs episode 10 in your head before running.

Describe one specific episode or moment that best shows your creature's personality.

Record the exploration rate and learning rate you used.

What actually happened? Describe how the bot moved and how its behavior changed across episodes.

Would you trust this bot in a real-world situation? What would you change about it and why?

Claude is ideal for developers, writers, professionals, and anyone needing advanced reasoning. It excels at coding, deep research, analyzing long documents, and automating workflows through native app connectors and terminal-based tools

What is Claude?

5th

Use what you observed to prompt Claude. It will generate a working HTML maze game. You can open it in any browser.

Turn it into a Game!

YOUR PROMPT FOR CLAUDE: copy, fill in the brackets, paste into claude.ai

Make me a simple platformer game where I control a creature and try to reach the exit.

The creature is called "[YOUR CREATURE NAME]".

The creature moves and reacts based on this personality: [CAUTIOUS / RECKLESS / EXPLORER / EFFICIENT / RANDOM].

Use these observed behaviors to shape the game:

1. It [paste one behavior]
2. It [paste another behavior]

Add this twist to the level: [paste your challenge question and turn it into a game mechanic].

Include:

1. Arrow keys or WASD controls
2. Platforms, obstacles, and an exit
3. An episode or level number
4. The player’s score
5. A restart button
6. A simple win or game-over screen

Create everything in one file using HTML, CSS, and JavaScript only.

5th

Turn it into a Game!

Paste your prompt into claude.ai and modify:

You can have a maze or platformer game

Behaviors: Runs fast, moves slowly, hides from enemies, collects everything, avoids water

Twist: Double jump, wall climb, temporary invisibility, slow-motion power, teleport short distances

Save it as [creature-name].html and open in your browser.

Does the bot behave the way you designed it? Watch a few episodes. If something's off, describe the bug to Claude and ask it to fix it.

Example