The robot empire is beginning to establish its dominance over the human race. Apparently, it thinks kicking butt at DOTA 2 is a good starting point.
Jokes aside, this is quite an incredible achievement. Developed by research lab OpenAI, this mechanical intelligence got its wits through an AI learning technique known as reinforcement learning. In reinforcement learning, the AI is thrown into a DOTA 2 game and given only the most basic of parameters to help it define what “victory” means. For example, the AI knows that death is bad, and winning is good. The rest is pure trial and error.
In order to give the AI system the greatest amount of possible experience, the program runs thousands of matches against itself, basically nonstop. This allows the AI to gain far more experience than a human ever could.
But this doesn’t mean the AI will have an easy time mastering the game. Unlike humans who might experiment on a game using something similar to trial and error but also reinforced by common sense and pattern recognition, this AI runs on pure trial and error. And thanks to a report by Endgadget, we know some of the numbers behind the AI’s calculations. At any frame in DOTA 2, a player can make one of about 1,000 different actions, many of which are the many degrees of possible movement. This is a far step up from the AI that once mastered chess, as the possible moves in a single turn of chess are somehwere around 35.
Add to this that the AI is controlling not one interface, but coordinating an entire team of five. And add to this that in order to understand what’s happening in the game, the AI must be able to track and respond to upwards of 20,000 individual values every frame. And these AI’s will not be learning the entire roster of 115 heroes, only 5. Savvy players might seek out heroes who are naturally strong against the AI’s picks, or heroes who are played so infrequently that the AI is unlikely to know how some of their mechanics work.
That’s not to mention the pure trial and error mentioned earlier: If a human player tries to run straight into enemy lines and dies, they might be more careful about that approach in the future. But if an AI sees a similar fate, the AI will mark that trial as a failure but will be unable to determine that it was their tactic itself that failed them. Next time, they might try running straight into enemy lines at an angle of one degree further to the left than their previous trial. Rinse and repeat.
Still, the AI does have a fair share of theoretical advantages. This includes a reaction time much faster than that of the average human, allowing for more coordinated plays. The AI’s single mind also eases coordination by eliminating the need for any communication between players, coaching, encouragement, or bad days. Once it gets up to speed, this AI will run games of DOTA just as consistently as DOTA’s own code boots up the game. It doesn’t have to account for the randomness of teammates or occasional flubs; once it really nails down the minutiae of evading a certain attack, it will do so every time, as long as that is the optimal course of action.
The killer robot will take on a team of top DOTA 2 players at upcoming DOTA 2 eSports event The International. Will human ingenuity be able to defeat this artificial force of destruction? Probably. Tune into the event on August 20th to find out.