Competition in different areas is normal for large companies, especially technological ones. Who will release a new product quicker? Whose product will be better? Who buys more companies? Of course, all of this is interesting but Google, Facebook, and Elon Musk have come up with a new kind of competition: battles in the gaming arena! So far not with each other. But maybe we should expect this?
A great battle – The beginning
Large technology companies now devote a lot of effort, resources, and time to develop and improve artificial intelligence and neural networks. What is the main essence of AI and neural networks? The correct imitation and independent reproduction of the logic of human behavior.
This is guides companies that make games. In some games with one player, for example, AI are the opponents of the player who try to kill or confuse you. In online games, AI are your rivals who first help you to understand the essence of the game and then they help to improve it. Although AI is almost always made weaker than the average player, they do have several difficulty levels. Except perhaps in StarCraft, but more on this later.
So we found out that AI in games corresponds to the general idea of using artificial intelligence — the imitation of human logic and actions. But before Google, Facebook, and Elon Musk started fighting artificial intelligence in digital games, Google decided to create an AI that could defeat a person in the most humane of human desktop games – Go.
It is believed that the ancient Chinese logic-strategic table game Go is uniquely played by each person. That is, roughly speaking, “play Go with me, and I’ll tell you who you are.” The game has two players, one uses white tiles and the other uses black on a square field of 19 by 19 lines. The goal is to capture as much space as possible with your color.
In 2014 Google acquired DeepMind, a British company that has long been involved in the development of artificial intelligence, specifically for the purpose of creating such an AI which is not only able to emulate the logic of a person playing Go, but also to surpass it. In 2015, the DQN system was jointly created, which was not informed in advance of the rules of the game, but it independently learned how to play 49 old arcade games and in some cases it even exceeded the results of humans.
A special algorithm, AlphaGo, was created to play Go. It was developed on the basis of combinations of the Monte Carlo method which use neural networks of policy and value. The neural network was trained on the basis of 160 thousand different matches in Go. The difficulty for the AI was that it is necessary to find a way out of 10 to 100 possible combinations, which far exceeds the combinations in chess. Prior to Google there had been several attempts to create AI for Go but they all played on an amateur level and Google wanted to defeat a real competitor.
And it happened. In January 2016, AlphaGo won against Fan Hui (three-time European champion). However, the public was not satisfied, believing that the European level of the game is not indicative. But Google was not going to stop there and challenged the legend, Lee Sedol, the best player in the last 10 years. The meeting of AlphaGo with the legend was scheduled for March 9, 2016. The meeting was also attended by two members of the Korean parliament and Eric Schmidt, the ex-head of Google. In total, 5 full-fledged matches were played.
Lee Sedol played black so he went first and AlphaGo was white. Despite the aggressive game play style of the legendary player, AlphaGo acted like a real human, starting to step back in just 12 turns. The system, according to experts, was looking for exits for itself. It was creating vulnerable moves and then turned them into its advantage. In the end, Lee Sedol admitted his defeat. After all 5 matches were finished, AlphaGo showed unprecedented results: out of 5 games artificial intelligence had won 4, refuting all skeptical attitudes of experts and expectations that an AI capable of beating a person in Go will only appear after decades of research. This was the first major victory of Google, on which the company decided not to stop.
But the two-time victory over mankind in Go did not satisfy Google. Almost immediately after the victory over Lee Sedol, Google announced a intergalactic war in the cyber-sport game of StarCraft!
StarCraft was not casually chosen, this strategy game is the standard in e-sports. The game requires very serious skills of quick reaction and prediction. Google admitted that it would be even more difficult than teaching AI to play Go. Although on the one hand the task is roughly similar: in StarCraft two players play each other and the task is to capture and defeat the enemy base. Only in this game the players do not have the opportunity to immediately see all the “battlefields”, as the players in Go can see. Thus, AI needed to predict what the enemy is doing at the other end of the map and this is a much more difficult task. Plus, AI needed to learn how to manipulate the mouse, use the camera intelligently, and effectively use the available resources.
In order to understand their actions, a separate API code was developed for AI which would allow it to “see” and “remember” its actions in a simplified scheme. It looks like this from the side:
And the company Blizzard is helping with this. They released an e-sports strategy. StarCaft is not the only game that has entered the sports discipline, but it is recognized by the gaming community as the most difficult to master.
The conquest of e-sports – Elon Musk
While Google is preparing its AI bot for a cybersport discipline, Elon Musk and his company along with Sam Altman of OpenAI created an AI that defeated one of the best players in Dota 2, Danil “Dendi” Ishtin, at the Dota 2 international championship in August of 2017.
Dota 2 is a multiplayer online battle arena. In the game there are 2 teams of 5 players who cooperate to destroy each other’s bases. Each member of the team plays for his own character which has its own characteristics, both in the use of different skills and in the strategy of the game.
In the rules of this match there was a condition that the winner would be recognized as the one who could make 2 kills or destroy the enemy’s tower. The first game was given to the AI quite easily, in just 10 minutes. Then the second match took place in which Danil surrendered after the first kill, admitting defeat. Elon Musk could not help but to rejoice at this event, which he reported on Twitter, not forgetting to mention the complexity of such an achievement in comparison with the victory in Go.
Indeed, as we said earlier in the case of Google and StarCraft, this task is difficult for artificial intelligence: the number of options in Go is infinitely large but limited. But a game based on a player’s behavior, which can not be predicted in advance, is much more difficult for AI.
This event was widely discussed in the gaming community, where many came to the conclusion that the AI had an unquestionable advantage over the human since its reactions are always accurate and thought out. The response in the game is much higher and more accurate than a person who uses a mouse for interactions.
Surprisingly, the AI initially did not lay out any strategy and OpenAI did not get help from experts. They just launched AI into the game and it started to learn on its own. At first it moved chaotically, but with each new movement it acquired more and more knowledge and within a couple of weeks it was able to reach the level of professionals, playing against himself.
Of course, it was just a 1 on 1 match, while a full-fledged match in Dota 2 is a game of 10 people. But OpenAI said they are not going to stop there and are already thinking about creating a team of 5 AI bots. This task will be more difficult. Since there is strategy used in a team game, it is necessary to take into account many more factors with a limited opportunity to see what the players of the enemy team are doing at the early stages of the game.
One of the main ideas of OpenAI is accessibility for all. Unlike commercial companies such as Google and Facebook, they are eager to share all the know-how they have for safe sharing of codes by everyone who wants to join and develop AI.
Intergalactic wars. Continuation
By creating AI for StarCraft, Facebook came into the mix in order to understand all the perspectives of the received data and SSP for its own purposes, including what the company made at the annual conference AIIDE (Artificial Intelligence and Interactive has the Digital Entertainment’s), during which there was a competition between AI’s in the game of StarCraft. Yes, unlike Google, Facebook started its development not from the version of AI against a person, but from AI against other AI.
In October of 2017, an AI lost to another AI which was developed by ordinary programmers who were fond of StarCraft and artificial intelligence in general. The detailed balance of power can be found here.
From Facebook, CherryPi took part. It was developed by Gabriel Sinnaev. Gabriel noted that CherryPi is just a “test pen” which will continue to be built on by the Facebook developers for more large-scale battles in StarCraft. By the way, Facebook sponsored this conference paying for the costs of equipment, on which there were thousands of games of bots.
This year’s victory was won by ZZZKBot, an AI created by Chris Coxe, who created it himself. But even this bot is unlikely to hold out against a live champion in StarCraft. Like none of the bots, according to the results of this tournament. There is no such AI that could have sufficient memory and planning skills to choose the right strategy, to have sufficient maneuvers, and the ability to predict the actions of a living enemy. But in general, the organizers of the annual bots tournaments for StarCraft note that the overall level of AI participants is significantly higher than in the previous year.
Facebook also published the largest data set to train the AI in the game of StarCraft. In total, 65 000 games are recorded in this data set, each of which contains exact data of all the players’ actions, army strength, movements on the map, and many other data. In addition, Blizzard’s collaboration with Google can help both Facebook and other AI developers.
While large companies are preparing for large-scale intergalactic battles, we can only marvel at how fast machines are being taught to perform various tasks. And to dream about what the future will be, where we will meet intelligent AI next time. Maybe in your refrigerator? And I wonder if Elon Musk will join the fight for the most clever AI in Starcraft or leave this achievement to his colleagues? Well, as they say, wait and see!