Musk's AI beats top eSports players, but this is not a breakthrough

(Original title: Musk's AI Robot defeats top human eSports player)

Qiu Zhili

At the Dota2 International Invitational, one of the world's largest e-sports events, Elon Musk’s non-profit artificial intelligence company OpenAI Robots overwhelmingly defeated top human professional player Dendi in a one-on-one match against Dota2.

For this achievement Elon Musk also expressed quite self-confidence, the first time said on a personal social platform, "OpenAI defeated the world's top eSports players for the first time, which is much harder than Go." Indeed In the process of the confrontation between the two, the OpenAI robot defeated the top human player in less than 10 minutes and won the first game. Such a strong technique allowed Dendi to withdraw from the second game and refused to play in the third game.

For a time, AI has captured eSports remarks and OpenAI has been deified. "This is an impressive achievement, but it is not an AI breakthrough." Denny Britz, a former member of the Google brain team, commented on the incident. "In fact, the over-hype of artificial intelligence progress is really dangerous. ."

AI still can't dominate the esports world

On the one hand, from the perspective of the game format, OpenAI is involved in a simpler 1V1 mode. In this mode, there is no match, and in most cases it is a frontal confrontation to understand the status of the opponent. "The non-perfect game problem is not serious among them, and the implementation path and state space of 1V1 is also much smaller." Tu Weiwei, a fourth paradigm senior machine learning architect, pointed out.

Tian Yuandong, a researcher at the Facebook artificial intelligence group, also knows that one-on-one is relatively easy, and the actual Dota25V5 game is a more challenging issue. 5V5 is a multi-agent problem. It requires a cooperation between agents to gain victory. It is a problem of non-perfect information game. Because of the “fog of war” in the game, players cannot know the full information of the game, and they are very large. The state space requires a very long execution path to get the game results.

In fact, OpenAI also hopes that in the "Dota2" season next year, it will be able to use the 5V5 approach again to take the challenge, but the 5V5 State AI is far more difficult than the 1V1 to defeat humans and is also more difficult than Go. Different from Musk's point of view, Tu Weiwei believes that OpenAI's defeat of eSports and Go's man-machine war is incomparable. This is two different tasks.

Currently AI technology is difficult to work well on two or more tasks. AlphaGo is designed for the 19x19 go task. The same version of AlphaGo cannot even be used on 19*19 and 20*20 chessboards at the same time. Work well. Currently, AlphaGo and OpenAI cannot exchange tasks with each other, which means that these artificial intelligence technologies currently do not have migration capabilities.

Tu Weiwei believes that the difficulty of Dota25V5 games is much greater than that of Go. "Dota25V5 is a multi-agent collaborative game, and Go is only a 1V1 game; the former has a number of states far greater than the number of states of Go and is a game of imperfect information, while Go is a perfect information game. Dota25V5 requires a longer-term strategy, often Thousands of operations have to be done before the results are achieved. In terms of technical contributions and social significance, OpenAI's robots are not comparable to AlphaGo."

At the same time, OpenAI chose Shadow Magic on the hero's choice. The hero itself does not have control skills. What is more needed is the player's precise judgments and subtle operations such as distance details. Tu Weiwei said: "This is precisely what humans are not good at, the machine is good at, the solo hero is often used by human players because humans are not good at such things, so that the hero stolo play can easily reflect the gap with other players So even if OpenAI wins humans in 1V1, it is far from solving the problem of 5V5."

The "immature" of OpenAI is also being verified. After the game, more than 50 players have defeated this deified AI. For example, some players deviated from the normal route by seduce local soldiers and let the other AI fall into a judgment to kill you. Shouhe's hesitation, and then win the game. “The AI ​​in Dota lacks strategic things, attacks, investigations, assistance, and other decisions,” said Wang Xi, a robotic machine learning scientist at Bamboo.

Staring at the AI ​​giant

In fact, it is not the OpenAI family who are interested in eSports. At the Wuzhen Weiqi Summit, DeepMind CEO Hasabis stated that interstellar hegemony will become the next big test for AI's progress. Followed by DeepMind and Blizzard in the "StarCraft 2" to carry out research experiments, and recently opened the training platform, while the game also attracted Facebook and Alibaba artificial intelligence team to join.

This real-time strategy game environment is complex and changeable, and tests AI's ability to understand a large amount of environmental information space, time, and data changes. At the same time, it also tests the collaboration, that is, the ability to co-operate with multiple units, buildings, and formations in the interfamiliar battle. It requires players to make high-level strategic decisions. This is why these giants use Starcraft 2 as a training target.

Although tens of millions of video game training AI procedures have been extracted for thousands of games, Blizzard said that the AI ​​is still in its infancy, and some pre-judgment tactics and collaboration capabilities can only be completed in small maps, and it costs a lot of money. The operation time can be used for the next step.

For the current e-sports industry, AI is more direct and realistic meaning of "generation training" function. Wang Hao, global partner of Frost & Sullivan and president of Greater China, believes: "On the one hand, AI can learn the experience of others and train new players. On the other hand, e-sports teaching also requires a lot of actual combat. AI is compared with human players. , will not be affected by emotions, time, response speed is also faster." After the game, Dendi said that in the process of watching the playback to learn something.

But the significance of giants developing artificial intelligence is far from playing chess or playing games, but is to better apply AI to real-world decisions. In the case of imperfect and complex information, how to make predictions and judgments and maximize the value of decision-making is the key to the development of AI.

In the previous demo war, the founder of innovation workshop Li Kaifu also explained to the First Financial Bureau the practical significance of artificial intelligence under “incomplete information”. “Most of the information in the world is not open, and it faces incomplete or misleading information. The ability of reasoning in the future will be able to resolve uncertainties in decision-making, diplomacy, business cooperation, and negotiation, and become a 'scholar' for humanity."

Education Projector

Educational projectors are mainly aimed at classroom teaching by school teachers; traditional projectors are not easy to carry. Due to the naughty students in school classrooms, projectors are not safe to place in classrooms and are easily damaged by students. The portability of educational micro-projectors makes up for teaching Vacancy, it is convenient for teachers to give lectures and only need to store the materials in the projector and show them to students for teaching, saving the trouble of textbooks and handwriting with pens and chalks. Its general weight will not exceed 0.2Kg, and some do not even need fan cooling or ultra-small silent fan cooling.

Educational projectors need to have:

1. Clear handwriting Educational projectors most often display text and images, at least a resolution of 1024*768 or more, and a contrast ratio of more than 3000 lumens, so that students in the back row can see the projection clearly

2. Short-focus large screen;

3. Accurate color;

4. Easy operation;

5. Features such as multiple connection methods.

education projector,education use projector,smart education projector,best education projectors,projector for education

Shenzhen Happybate Trading Co.,LTD , https://www.happybateprojectors.com