Top 100 Russian high school students have been building algorithms to predict the outcome of Dota 2 matches as part of a data hackathon sponsored by Russia’s largest financial institution Sberbank. GOSU.AI provided the contest with the necessary data, around 30,000 matches in this case.
The participants in the data science competition ‘Sberbank AI Academy’ built algorithms using machine learning techniques. The goal of the task was simple. A given replay of a random game is stopped at a certain point and the students need to calculate the probability of a team to win. Each replay provides plenty of data to be used and taken into consideration. A lot of small details, including chat logs also fell into that category. Most impressive indeed, with a threshold of 0.75, meaning that if Radiant/Dire had a 75% chance to win, they’d call it and if below that, it’s considered a loss, they managed to get 66% of the given matches correctly. Were the threshold 0.5 instead, the percentage would raise to 90.
You can find further details on the official site of the competition, with the added option to download all the results. GOSU.AI mentioned they’ll soon release data from more science fairs with different tasks on popular eSports titles like CS:GO, PUBG and Dota 2.