Well, competing in major eSports championships in the future can become a nightmare if machines can one day register their own teams. An artificial intelligence called OpenAI Five managed to beat humans in a match against an amateur group of Defense of the Ancients 2 (DotA 2). For that, she needed five neural networks working simultaneously.

But don’t think that computers don’t need to train, just like meat and bone sets do. To achieve this result, OpenAI Five must simulate 180 years of rounds against itself daily. That’s right that you read, 180 years in a day, to learn from your own mistakes and successes.

Obviously, respectful hardware is also required for this. The system uses 256 GPUs and 128 thousand processing cores, which use Proximal Policy Optimization, a collection of codes that improves learning with algorithms. With this schedule, OpenAI Five is able to determine the objectives more efficiently.

“Using a separate network of long-term memory for each hero and no human data, OpenAI Five learns recognizable strategies. This indicates that reinforcement learning can yield long-term planning on a large, but achievable scale – without fundamental advances, contrary to our own expectations when starting the project ”, explained a project representative.

How it works?

Well, the OpenAI Five was successful on seven fundamental points for its victory to be realized. The first was to perform a good team work, taking advantage of the characteristics of each character and efficiently neutralizing the main pieces of the opponent. The second was the prediction of values, which made it possible to anticipate the maneuvers and their impact on the match.

Then she had to learn to search the forest, to find enemies who tried to hide behind the trees. The machine also needed to reposition itself quickly and strategically, to the point of being able to surprise its opponents. Another key attitude for victory was to focus on the multiple skills that cost the most to ensure a more valuable enemy death. In second to last place was the chase, in which the machine had to understand how to anticipate the movement of the other players to prevent them from escaping.

And finally, OpenAI figured out how to create a distraction, like sacrificing a hero so that your team could achieve the goal. All of this is duly illustrated and commented on in the video below:

Now, OpenAI Five continues to train with a more ambitious goal: to defeat a professional DotA 2 team.