1From keystrokes to 48 numbers
Pick a game or edit the log. The numbers are computed with the same code as our notebook.
s is a selection, Base the main building, hotkey32 is slot 3, action 2
(0 created, 1 updated, 2 used), and t5 means 5 seconds of game time have passed.
Time windows
Each time marker closes a window. A window counts for at least 1 second.
Three phases
The 48 features
Each count is divided by the seconds of its phase. Click a cell to see the calculation.
Compare with the whole game
The notebook also printed the features of this game on its full log, which isn't public. pandas only showed the first and last columns, so there are 16 of the 45 rates.
2Which features the forest uses
Two ways to measure it, on the random forest v1. The notebook only kept the plots, so I read the values back from them (about ±0.0001).
How to read it
Impurity decrease: how much the splits on a feature clean up the trees. The 48 values add up to 1. Permutation: how much the accuracy drops when only this column is shuffled, averaged over 30 shuffles. The line shows the spread.
3Seven models, same split
All trained on 80% of the 3,052 training games and tested on the same other 20%, as printed in the notebooks. The axis starts at 85%.
Selected model
About the data
StarCraft II is a real-time strategy game. Good players use hotkeys a lot: they bind groups of units or buildings to the keys 0 to 9 and switch between them dozens of times per minute. We had 3,052 games from 200 players, each one written as a single line of tokens, from a few tokens to more than 10,000.
A model needs the same columns for every game, so we cut each game at 25 minutes and split it in three phases: start (first minute), early (up to 6 minutes) and mid (up to 25). In each phase we count, per second, the presses of each hotkey slot, the hotkeys created and used, and the selections. That's 15 rates for 3 phases, plus 3 columns for the race.
The real games were on Kaggle, behind the class competition, so they aren't here. The logs you can load above are the only parts our notebooks printed: the first seconds of ten games, and the example from the competition page. It's already enough to see a habit: training games 0, 2 and 4 are all player 53, and all three create hotkey 3 in the first seconds.