Intro to Machine Learning final project

Who Wins at Uno?

We simulated 99,989 Uno games under five sets of house rules, then trained a model to predict the winner while the game is still being played.

Watch a game
Kareem Bismar and Russell Wilner
7
+2
4
?
99,989
simulated games
50.5–55.3%
of winners picked by the model, depending on the rules
45–47%
picked by betting on the fewest cards
25%
picked by guessing at random

The question

Can we predict who wins before the game ends, and do house rules make the winner easier or harder to predict?

The method

Four bots play each game with a random, greedy or strategic style. We snapshot every game at five points and compare logistic regression, random forest and gradient boosting against a simple fewest-cards rule.

What we found

Gradient boosting beat the simple rule under every rule set. Predictions get much better late in the game. Stacking made games the most predictable and draw-until-play the least.