Project idea · ungraded

Grid Game Agent

Design a small grid game of your own — collect keys and reach the door, avoid moving guards, or push a box onto a target — as a Gymnasium environment, then train a Q-learning agent to play it and watch it improve.

You will use

Milestones

Build the smallest working version first. Ticks are saved in this browser only.

01Define the rules, the state encoding and the rewards on paper; keep the state space small.
02Implement the game as a gymnasium.Env and pass check_env.
03Print the grid as text after each step so you can play it yourself with the keyboard.
04Train Q-learning and plot the success rate and average return per 100 episodes.
05Compare Q-learning with SARSA and explain any difference in the learned behaviour.

Stretch goals

Make the game stochastic (slippery moves, random guards) and compare again.
Replace the table with a small neural network (deep Q-learning) for a larger grid.
Record and replay an episode as an animated GIF.

Data

No data is needed: the agent learns from playing your game. See the FrozenLake documentation for a text-based environment to imitate.

Run it yourself

pip install gymnasium numpy matplotlib
python train_agent.py