Detailed analysis of the Dota 2 match between OpenAI and people in the 5x5 format. People lost

    Yesterday, August 5, in San Francisco, there was a show match between people and OpenAI bots in Dota 2 discipline. Back in 2017, in the framework of The International 2017 show matches, people fought OpenAI in the 1x1 mirror mid format and with a number of restrictions in favor of bots (a ban on the use of a number of items and a mechanic), which resulted in the defeat of professional mid-players.

    Since Dota 2 is an extremely versatile discipline and difficult to master, the meeting between people and AI was again held with a number of restrictions, which, however, did not have a very radical effect on the gameplay:

    • пул из 18 героев в режиме Random Draft (Axe, Crystal Maiden, Death Prophet, Earthshaker, Gyrocopter, Lich, Lion, Necrophos, Queen of Pain, Razor, Riki, Shadow Fiend, Slark, Sniper, Sven, Tidehunter, Viper, или Witch Doctor);
    • без Divine Rapier, Bottle;
    • без подконтрольных существ и иллюзий;
    • матч с пятью курьерами (ими нельзя скаутить и танковать);
    • без использования скана.

    The most serious limitation: an extremely small pool of heroes for both parties. Now in Dota 2, there are 115 characters with different abilities and mechanics of their use. OpenAI can cope with only 18 of them so far. The meeting was as close as possible to the “real” conditions and was held in 5x5 format. Ordinary people who had once been involved in eSports, but now are not eSports players, played against AI. The only concession for people was that the reaction of the bots was limited to 200 ms in order to avoid situations with instant “pressing” of the buttons. Bottom line: the AI ​​team won 2-0 on cards. It was possible to win from OpenAI only after the heroes for AI chose the auditorium (Slark, Sven, Ax, Riki and Queen of Pain), according to OpenAI estimates, the chance to win with such a draft was only 2.9%. In addition, before the main match, ordinary guests of the event could play with bots, and in these meetings the dominance of the AI ​​was even more obvious, which is impressive.

    Why is it so important

    OpenAI is a non-profit project, behind which, however, there are quite well-known and influential people. The project was founded in 2015 by Ilon Mask, Sam Altman , and the sponsors were Peter Thiel and Jessica Livingston . Dota 2 is not the only activity of the OpenAI project team. If you read the section "Research" on the official website of the project, you can see that OpenAI is also actively working towards creating AI for, for example, robotics.

    But if the founders of the project are such influential people, then why did the choice fall on a computer game? Learning AI to work in dynamically changing conditions is not a trivial task. The current models of AI and autopilots operate exclusively under rigidly defined conditions, and their training requires millions of different combinations of events and factors so that the bot does not just act on pre-defined patterns, but begins to analyze what is happening and make decisions based on the context of the situation. Choosing Dota 2 as a platform for teaching AI and creating bots for this discipline was useful from many points of view:

    • eSports has high media coverage;
    • it is a safe platform for practicing AI teaching methods;
    • the huge variation in game mechanics Dota 2 provides an almost unlimited number of scenarios for learning;
    • open access to live game records provides an inexhaustible supply of data for analysis.

    The last item on this improvised list is probably the most important one. One of the problems of all AI training systems is the creation of a relevant sample on which to train the system. Live players generate absolutely free millions of game entries for free on a daily basis, and researchers can only choose matches with a relevant set of characters. And the result is impressive.

    What happened on August 5

    Most often, bots in computer games are either stupid scripted dummies, or invincible monsters with an inhuman reaction. As an example - “insane” -bots for CS 1.6, which were killed in the head with a shotgun at a distance of 50+ meters in a leap with a turn . To achieve realistic, but at the same time effective behavior from gaming AI, developers have been failing for already a decade, which is especially evident in open world games. Actually, this was the reason for the popularity of online disciplines.

    After the show matches in 2017 in the “1x1 mirror” format, the OpenAI team stated that the next goal is full 5x5 games. And they almost managed to achieve the intended goal: it would be silly to expect bots to control and analyze the synergy of all 115 characters in Captain Mode, but already 18 heroes are a lot. Here is the calculation of the options for the peak 5x5 (10 characters) for 18 heroes:

    C 10 18 = 18 !10 ! ( 18 - 10 ) ! =18!10 ! 8 ! =111213. . . 1617181 2 3 . . . 6 7 8 =43758

    But such a number of options we get for 115 characters:

    C 10 115 = 115 !10 ! ( 115 - 10 ) ! =115!10!105!=106107108...113114115123...8910=74540394223878

    Yes, 74.5 trillion combinations! Obviously, in order to cope with such diversity and include all 115 characters in the OpenAI pool, a full-fledged self-learning AI is required, which we do not yet have. At the same time, the calculations do not take into account the abilities of characters (from 4 to 6 “buttons” on a character, without Invoker), types of abilities (active, passive, directional, non-directional, AoE, aimed at a point), dozens of objects, active effects of objects, modifiers And so on and so forth.

    But even for our mini-pool of 18 characters, the work was tremendous.

    Below you can evaluate how OpenAI bots behave in games against people:

    How bots differed and did not differ from people

    For those who are too lazy to watch 5+ hours of recording from Twitch, we will briefly tell you how the behavior of the OpenAI Five bots team was remarkable against ordinary people.

    Immediately it should be said that the AI ​​was thinking at the draft character stage, that is, he was counting the chances of winning one or another combination of characters based on the incomplete data he has. At the same time, OpenAI reported the predicted percentage of its victory based on the draft. Sometimes, according to AI, it reached 95%.

    The whole OpenAI Five game can be described by such a popular term with the community as “choked”. Yes, according to the sensations, the bots actually “choked” certain lines, while without any remorse or fear they came under the towers in the first minutes of the match in order to send the human player to rest under the fountain at the base. Some points can be seen on the thread below (the title of the video does not correspond to reality):

    In addition to fearlessness (bordering on arrogance) and organized work at an early stage of the game, the bots were noted by Dota 2’s own understanding of the mechanics. The goal of the game is to break the main structure of the enemy, on which the bots were focused. Thus, 3-4 AI characters consistently broke the towers of human opponents, masterfully “tanking” the tower in turn, in order to preserve their own creeps and quickly destroy the structure. In addition, bots were noted by the fact that non-standard wards were used, which also tanked towers for them. As a result, in one of the games with guests of the event, bots entered the enemy’s base already at the 9th minute (climbing up to the enemy’s base, even at 15–18 minutes, is considered to be exclusively early timing).

    Also, bots effectively redistributed the experience between themselves, avoiding drawdown of characters 4-5 positions, distinguished themselves in calculating the damage (when the last strike was applied, the bot did not continue to catch up with the target, clearly realizing that there was enough damage to kill) and actively used some items.

    The attentive viewer will notice on the video and making decisions based on incomplete data (using AoE skills in the “fog of war”), making non-standard decisions (successful use of the teleportation scroll instead of trying to escape on your own two feet), entering the rear with blocking the enemy in the forest and coordinated focus of the most dangerous human opponents.

    In order to justify humanity, we can say that human players showed not too high level of the game, were limited in terms of the choice of characters and made many mistakes, but such progress by the AI ​​in just 1 year of development can not fail to impress.

    It is likely that we will see another show match, this time at The International 2018 (will go August 20-25) against professionals who are clearly stronger than AI. However, it is quite possible, OpenAI Five will be able to surprise us with something.

    Also popular now: