Earlier this year we reported on a company called MoreMMR using artificial intelligence technology to predict matches in The International Dota 2 tournament, and we recently caught up with them to take a deeper look into what happened and where the idea came from.
Where did MoreMMR come from?
After successfully exiting the previous business, Maxim Dreval, the founder and CEO of L2P ltd. was looking for an attractive business area, that's when esports appeared. The gaming industry and e-sports in particular is evolving with a huge pace. We built the core part of our team in a matter of days, everybody was excited and full of ideas, that's where it all began.
One of games that attracted everybody's attention with a huge prize pool was Dota2, announcing $1M [USD] prize fund for The International tournament in 2011. Since then, prize pools reached $25M and the game gained tremendous popularity. 3 of [the] 5 founders team have played Dota2 so choosing it as a first game was obvious decision.
But Dota2 is very variative game, with lots of factors influencing the probability to win. In order to improve in it, you need guidance and we provide it in many ways.
Where did the idea for the AI come from?
For the last 10 years, machine learning has become very popular; it is used everywhere and our platform is not an exception. As mentioned before, Dota2 is very variative, so we decided to make a model that would analyse which indicators in match are most valuable in terms of winning probability and highlight the mistakes for the user that are based on those crucial factors. So, the post-match analysis tool was born.
How does the AI technology work?
We have already developed a system that recommends you the ideal build in the game. Our service analyses mistakes in the game and now we turn from expert mistakes to mistakes founded by AI. (e.g. we have a model, that checks if it necessary to buy a bkb or linken sphere in a certain match or no). We have a huge DWH with logs of historical matches. Depending on the task, we form the necessary data set and train various models on it (e.g. any kind of decisions tree algorithm). Sometimes, when we work with categorical features, we use embeddings techniques, to reduce dimension of feature space.
How did the AI prepare for The International?
The model was based on the matches played by pro players. In fact, we uploaded around 2 000 matches from official tournaments. The model was learning to predict a winning probability, based on several parameters of the match. The model took into account both individual skills of players and team interactions to improve the accuracy of the prediction. We use neural net to encode categorical features (e.g. items, collected by player). This embedding technique during tests showed itself much better than a standard one-hot encoding. And with this approach dimension of space doesn't grow significantly.
Finally, all features are going to decision tree to make a prediction.
How was this used in The International?
We conducted a contest on our platform where users could make their predictions about the winners of each match played on TI8. The point of the contest was "the man against machine" because - along with users' predictions - our prediction model was fighting for the prize too (we would never take away the prize from our users).
What were the results like?
This International had very unexpected plot twists throughout the whole event. The model took the 12th place out of 6000 participants, which is good, considering what unpredictable matches were played in the tournament. But, unfortunately for us, the model put its bet on PSG.LGD in the finals, although we are very happy for OG and their excellent performance. It was really the coolest finals of recent TIs.
What lies ahead for the future?
The next step is to develop a model that can predict not only the winner of the match but also more complex events - the first blood in the game, the first tower down, and so on. We already have some developments and plan to release them by next Major. As for the users' post-match analysis, we will continue to develop this area, adding more deep mistakes, skills recommendation, playstyle analysis and more.