cool hit counter AI takedown of Dota 2! 5v5 Group Battle Beats Humans for the First Time, Already Exceeds 90% of Human Players_Intefrankly

AI takedown of Dota 2! 5v5 Group Battle Beats Humans for the First Time, Already Exceeds 90% of Human Players


When Deep Blue and AlphaGo killed all the strongest players in chess and Go in solitary defeat, many felt that humans were no longer able to resist the challenge of AI in decision-making games. However, at least in eSports, and especially Multiplayer online tactical competitive games (MOBA, e.g. Dota 2) The AI has yet to overrun a human fortress in this kind of strategy game that requires teamwork.

On June 25, OpenAI, a non-profit AI research company founded by Elon Musk and Sam Altman, announced that they had made an important breakthrough in the field of. The OpenAI Five system, trained through reinforcement learning, defeated amateurs in a 5v5 team battle in Dota 2. Next, they aim to challenge top professional teams at The International, the most important event in the Dota 2 community this year.

Source | OpenAI, the Verge, etc.

Compile | Wufei

Video credit: OpenAI

Challenging machine collaboration challenges

Greg Brockman, CTO of OpenAI, called it a landmark breakthrough: "This work is significant in real-world applications, when you can model a problem, you can apply it to larger scenarios without barriers. "

Last August, OpenAI took a stab at Dota 2. They designed the 1v1 bot system to beat the top pro's in 1v1 matchups. Of course, 1v1 matchmaking is much easier to implement than group battles. As a MOBA, Dota 2's difficulty lies in the teamwork: each side consists of five players, each controlling a "hero", or in-game character. On a map, teams win by dividing up the work to push back the other team's base. but (not) Teamwork, too, is an unprecedented challenge for AI systems.

Some of the members of the OpenAI team. Last year, it was with this laptop that the AI beat the top pros in 1v1 matchmaking.

The reason why artificial intelligence has struggled to overcome MOBA games such as Dota 2, in addition to the teamwork required between systems, is that such games require long-term decision-making - in 45 minutes of gameplay, players need to perform tens of thousands of actions, and the impact of some of these decisions will be felt throughout the game; there is a limit to the amount of information each player can see, especially from the opponent, compared to the information transparency of board games; in games, systems need to process 20,000 data points in a split second and make choices from 1,000 different actions, far more than in a chess match.

Faced with this set of challenges, building on the 1v1 bot, OpenAI researchers have developedOpenAI FiveAlgorithm. The researchers used reinforcement learning methods to allow the AI system to learn by constantly playing against itself in a virtual environment. Their daily volume of self matchups is staggering, the equivalent of 180 years of play. That means that the AI trains in a day, which is tens of times the amount the most experienced pro player trains in a lifetime.

The training process used 256 GPUs and 128,000 CPU cores, a significant improvement over last year's 1v1 bot system. Each hero uses his or her own long- and short-term memory network, and no human data is used for the entire process. "For the first few hours, the AI will just randomly wander around the map. But gradually, it was able to master some basic functions. "Brockman said.

Greg Brockman

Defeating the amateurs

The AI system has traits that human players would envy. It's more responsive. that takes only 80 milliseconds, faster than human players. Can perform 150~170 operations per minute that is on par with the top human pros and doesn't miss a click. The AI is also superior in that they have ready and accurate access to important information such as distance between characters, equipment bars, and hero health, and apply this data to choose the best strategy. In contrast, human players can only check manually, or judge from experience and instinct.

Since May of this year, OpenAI Five has competed in 5v5 team battles against five amateur and semi-pro teams of varying levels, and the research team was pleased with the results. Even against well-trained semi-pro teams, the AI won two out of three matches. And in a match against the OpenAI staff team, the AI even came out on top. OpenAI Five has an average ladder score of over 4,200.

It is worth mentioning that OpenAI Five showed a strong sense of teamwork in the real world. "They know how to sacrifice a line or a hero for the ultimate victory." Brockman said. The AI also received praise from its human teammates for its performance. In one round of testing, a human player joined the AI team. He says that his four AI teammates have given him plenty of support: 'Whatever I want, these machines can make it happen. "

After winning the series of tests, the research team set their sights on a more formidable opponent. They plan to continue optimizing their system and compete against top professional teams at The International this July 28th.

Broader application scenarios

While OpenAI Five has already had good results in 5v5 live matches, it is important to emphasize that its applicability is currently more limited. Dota 2 players know that they need to choose five heroes from 115, but at the moment OpenAI Five can't do that, it can only play with five selected heroes [Plague Mage (Necrophos), Sniper (Sniper), Underworld Subterranean Dragon (Viper), Crystal Maiden (Crystal Maiden) and Witch (Lich)]. No need to feel sympathy for the AI's choice of heroes, as human players are also limited to the same characters; in addition, the "Man vs. Machine" is limited in its rules: no eyes, no flesh mountains, no stealth, no summoning and no illusions ...... If you are puzzled by these terms, remember that these are mostly more difficult operations for decision making.

While the study leaves something to be desired, in the OpenAI team's opinion, the research has implications far beyond the game itself. For example, in the real world, the AI needs to give real-time feedback on what is happening at any moment, and such abilities are one of the keys to AI winning in Dota 2 matches, but cannot be trained in turn-based chess matches. In addition, Dota 2 requires the AI to make decisions without access to complete information, which is very similar to a large number of practical application scenarios. Being able to master the Dota 2 game means that AI will likely play a bigger role in real-life scenarios such as urban transportation systems and logistics systems.

Reference link.

https://blog.openai.com/openai-five/

https://www.theverge.com/2018/6/25/17492918/openai-dota-2-bot-ai-five-5v5-matches

https://www.newscientist.com/article/2172612-ai-trained-on-3500-years-of-games-finally-beats-humans-at-dota-2/

https://www.engadget.com/2018/06/25/openai-bots-dota-2-the-invitational/


Recommended>>
1、Seamless scrolling case simple implementation code
2、Golang Language Community Build your own standalone development system environment
3、Why cant browsers cross domains
4、Ali interview questions summary
5、Code to wait a few seconds after tapping a button to stop in a rotating image and then play it automatically

    已推荐到看一看 和朋友分享想法
    最多200字,当前共 发送

    已发送

    朋友将在看一看看到

    确定
    分享你的想法...
    取消

    分享想法到看一看

    确定
    最多200字,当前共

    发送中

    网络异常,请稍后重试

    微信扫一扫
    关注该公众号