OpenAI bot crushes the champions of Dota 2, and everyone can play against him

In recent years, OpenAI, a start-up whose mission is to ensure that "artificial intelligence general benefits all humankind", has developed a robot based on the machine learning Dota 2, the biggest game in the universe. Starting from a very small version of the full game, the bot has been developed over the years by playing millions of matches against itself, learning not only to play the game at five, but also to win, regularly.

We could follow the development of the bot during several games, each using a more complete version of a game and more skillful opponents. This culminated in what should be the last match of the show this weekend, when OpenAI Five played a draw against OG, the team that won the biggest competition of all the sports esport l 39; last year, The International.

OpenAI is subject to some handicaps in order to keep things interesting. Each of his five AI players runs an identical version of the bot software, without any communication between them: they are five independent players who are very similar but who have no direct way to coordinate their actions. The reaction time of OpenAI is artificially slowed to make sure that the game is not simply a showcase of superhuman reflexes. And the bot still does not use the full version of the game: only a limited selection of heroes is available, and objects creating controllable minions or illusions are forbidden because it is thought that they would be able to manage his henchmen more effectively than using them. any human could.

The games can be viewed here. The first match seemed homogeneous until about 19 minutes. Humans had a small gold advantage, but robots had better territorial control. The bots came out winners in a team fight, killing three human players while losing only one. The game still looked like a knife, but the bots were not in agreement: they announced that they had a 95% chance of winning and, after making that statement, they immediately used their numerical advantage to inflict heavy damage to the human base. . This further strengthened their territorial control and also gave them a significant lead in gold.

This put the humans on the back foot and, although they managed to shoot the game for another 20 minutes, they were not able to overtake the robots ahead, thus giving at OpenAI a 1-0 advantage.

In the second match, things were not even close. the robots quickly took advantage and pierced the human base in less than 15 minutes. They won five minutes later.

Overall, OpenAI was a dominant performance: a 2-0 win against an established human team accustomed to playing with each other at the highest level of the game. This performance was by far the strongest. OpenAI over the years.

The coordination of robots is strange: although they can not communicate, the five computer-controlled players think the same way. If we think it's a good opportunity to attack a human player, the other four of them will think the same way and will participate in the attack. This gives the appearance great coordination in team battles – coordination with precision and rigor that human teams can not match.

But OpenAI seems to be beatable. It has certain weaknesses, even if they are surprising, – it's not great to score the last shots, the fatal blows to the computer-controlled units that are used to accumulate gold in game. This gives humans the opportunity to get a early gold advantage. Robots have also struggled to counter the invisibility of the human side. They also seemed ill suited to certain spells of some heroes, especially the Earthshaker Crack, a spell that temporarily creates an impassable barrier on the map. Humans have been effective in using this to trap bot players and restrict their movements, and this seemed to confuse OpenAI.

The behavior of robots is also a lesson from the big gap between this type of machine learning system and a general artificial intelligence. Although AI Five is clearly effective in winning games, it is not clear how to play. Dota 2. The game's players use a technique called "shoot" to redirect the stream of computer-controlled minions from their side (known as creeps in Dota 2) as a way to deprive the enemy team of gold and experience. Human players may recognize that this has happened because goose bumps do not show up when they are supposed to. Human players have a mental model of the whole game, an understanding of its rules and can recognize that something is wrong; they can reason about where goose bumps had to go and interfere with the attraction. On the other hand, the computer is wandering aimlessly at this scenario.

No pulling

OpenAI does not seem to have used the shot-on-goal technique in millions of games against itself and has never learned to play against it. So, when a human team starts shooting, the bot does not recognize the situation and does not really know what to do. It can not reason about what the game should be like, or about the unexpected behavior of the game. All the bot can do is look for the patterns it recognizes and choose the action most likely to produce the best result; give him a motive that he can not recognize and its performance is deteriorating.

Until now, the OpenAI bot has been restricted. some pros and streamers have had the opportunity to play against it, and it has also been available against some live events. But for a few days, it changes: Dota 2 players can register here to play against the bot – or with him – for a period of three days. Unfortunately, this public period does not seem to give rise to a new, improved bot: beating the goal of a top-notch human team was the goal set by OpenAI for its bot and, with that accomplished, the experience seems to be over.