Chinanews client, Beijing, July 15 (Reporter Song Yusheng) "The overall feeling is that AI is too powerful, and they can always react differently from ordinary people..."
After losing 1:3 to artificial intelligence, some professional e-sports players of Glory of the King said about their feelings during the war.
Since then, more netizens have bluntly stated on social media, "Humans can't even play games against AI" and "Humans are once again defeated"...
So, let the artificial intelligence to play "King of Glory" is to "abuse food"?
A reporter from Chinanews.com recently chatted with the person in charge of the AI model.
The scene of the game.
Photo courtesy of the interviewee
3:1 "It belongs to normal play"
A few days ago, at the World Artificial Intelligence Conference site, a game AI named "Jue Wu" played an exhibition match with five players from the King of Glory Professional League.
It is reported that the artificial intelligence has successively performed operations such as "squatting grass", "pulling output back and forth" and "assisting team opening" in the competition, and has a professional sense of teamwork.
In this five-game three-win event, the artificial intelligence defeated five professional players with a score of 3:1.
Although the outcome of the game has become a hot topic of discussion among netizens who are following this game, Qiu Fuhao, the head of the AI model of King Jue Wu, told reporters that the team had no expectations for the outcome of the game.
As R&D personnel, they hope that artificial intelligence can grasp the opportunity to compete with high-level players, verify their abilities and continue to grow, although such results are "normal play."
The scene of the game.
Photo courtesy of the interviewee
In addition, many netizens speculate that, compared to the reaction speed of human professional players, artificial intelligence may have an "absolute advantage" in operation.
However, Qiu Fuhao clearly stated to reporters that in terms of numerical value and operation, King Juewu AI does not have unique advantages, and the objective conditions in the game are the same.
"In terms of the hero's own state parameters, AI does not have additional bonuses, which are the same as human players; AI is consistent with human players in terms of field of view observation, and AI can also not see invisible units in the fog of war; AI The operational response of the game is also objectively restricted, and its response distribution and average are similar to those of professional KPL players. According to actual observations, humans playing Luna and other heroes will be better than AI."
The scene of the game.
Photo courtesy of the interviewee
AI games per day ≈ 440 years of humanity
Since objective conditions are no different, why can artificial intelligence defeat human professional players?
The answer given by Qiu Fuhao is the amount of training.
"The advantage of AI lies in its huge amount of training. The number of games in a day is approximately equal to 440 years of humans."
In the words of the R&D team, since the start of the project in 2017, Wang Zhe Juewu AI has gone from imitating human supervised learning to self-playing reinforcement learning. After many iterations and overcoming many problems, it has gradually evolved into an all-hero career. The "complete body" of the e-sports level.
Qiu Fuhao introduced that in 2018, King's Juewu AI was only at the level of top amateur players; by 2019, King's Juewu AI no longer needs to imitate human data, but by competing against itself to further improve the level of micromanagement and the overall view. It has reached the level of professional e-sports of the glory of the king; last year, it learned more heroes to play, and used 40 heroes to accept the challenge of the players for the first time;性optimization.
Data map: 2016 Go man-machine battle.
A few years ago, AlphaGO shocked the world when it defeated a human professional Go player.
Compared with Go, "King of Glory" is an incomplete information game scene that requires multiple AIs to collaborate to complete tasks, and long-term decisions need to be made in a complex and continuous decision space.
In other words, teaching artificial intelligence to play "Glory of the King" is more difficult than allowing artificial intelligence to play Go.
Qiu Fuhao said that the action space of Go is as high as 10 to the 172th power, while the operation possibilities of AI in a game of Glory of the King are as much as 10 to the 20000 power.
The scene of the game.
Photo courtesy of the interviewee
Let AI play games is not the end
But, is it only for artificial intelligence to defeat humans in the game to carry out such complex R&D and training?
Qiu Fuhao didn't think so.
Because artificial intelligence is ultimately to serve mankind.
He told reporters that the reason why the team developed this is firstly because "this is a multi-player collaborative game scene. Its design is highly complex and challenging, which meets the research needs of high-level AI+ games."
"From the key events in the development of AI in recent years, we can see that the next milestone of AI is likely to be born in a complex strategy game."
The purpose of artificial intelligence for complex game training is to help humans solve real-life problems.
The scene of the game.
Photo courtesy of the interviewee
Qiu Fuhao said that AI learning general perception and decision-making capabilities in a virtual environment may play a greater role in the real world in the future.
Taking the study of unmanned vehicles as an example, AI learns the external environment, perceives and expresses road environment information, and then makes driving decisions; another example is the simulation learning of robots in the virtual world, and then reacts to unknown changes in the environment in the real world.
He believes that in the future, AI research will cover more scenarios.
For many researchers and developers, there are still significant practical difficulties in multi-agent technology research, including the uncertainty of the environment, the limitations of information acquisition, the consistency of individual goals and global goals, and the impact of high computing power. Requirements.
In Qiu Fuhao's view, "AI+Game" research will be a key step for the team to overcome the ultimate AI research problem-General Artificial Intelligence (AGI).
AGI stands for the development of AI that can execute a variety of complex commands in a general-purpose system, reaching or surpassing human level.
"The experience, methods, and conclusions in the middle are expected to have a more profound impact in a wide range of fields, such as medical care, manufacturing, driverless, agriculture, and smart city management." (End)