Reporting by XinZhiyuan
OpenAI stood up for 45 minutes, but still lost!
After losing the first game yesterday, in today's Ti8 exhibition match, the OpenAI Five team once again played against humans.
The winners were five former pros from China who beat OpenAI Five after a 45-minute battle.
The five contestants are:
Zhang Ning (xiao8), dota, DOTA2 professional player, game ID: xiao8, played for the NEWBEE team, joined the LGD team in 2014 as the leader of the LGD team.
Luo Fei chi (430), a Dota/Dota2 pro, played for IG Esports Club as the Number 2 position, and was nicknamed "Pianist" because of the fact that the Carl he used was at the top of the world.
Wang Zhaohui (sansheng), DotA pro, TI4 World Champion member. ID: sansheng, position: fifth position support, has played for teams such as tongfu (tongfu), known as dog brother.
Xu Zhilei (burning), the world-famous carry player, is considered to be one of the best and most experienced carry players in the world, and has won many carry awards and world championships over the years.
White sail (rotk), as the 3rd position, and as the BP. Good at Bat Knight, Dark Sage, Tide Hunter. He led team TYLOO to the DMT Championship and was a member of the 2011 DK team.9 championship.
(Imaged on Dota Cinema Twitter)
Compared with yesterday, the lineups of the two sides of the match were simply exchanged, namely:
OpenAI: Ryan, Plague Mage, Witch Doctor, Sniper, Axe King;
Human players: Tidal Hunter, Lover Helicopter, Crystal Virgin, Death Prophet, Lich;
Based on the results of yesterday's exhibition match, the OpenAI team said that they would make further modifications and debugging.
Therefore, in the early and middle stages of today's game, it can be seen that the mistakes of AI players are not as many as yesterday; but in the later stage, the OpenAI Five team frequently buggies, living like a few headless flies, which eventually led to the failure of the game.
From the overall performance point of view, OpenAI still reflects the advantages of human beings in terms of response speed, calculation accuracy, operation, etc. Team strategy, flexibility, and response to unexpected situations still need to be improved.
Below is a recap of the tournament.
The human Jedi strikes back, and several waves of GANK defeat Bug's OpenAI
As soon as the game began, human players hugged tightly, OpenAI gave a blood without suspense, however, AI did not show weakness, immediately counterattacked, and finally the two sides ended the opening with "2 for 2".
In the early development process, AI pays more attention to team battles, while human players focus more on their own development, resulting in human players lagging behind in the number of heads in the early stage.
Although the situation of AI insertion today is somewhat better than yesterday (at least there is no excessive random insertion), there are still phenomena such as "fake eye insertion next to the real eye" and "installing a camera at home":
In the first wave of wonderful GANK, human players wanted to push the tower in the middle of the road, but they were counterattacked by the AI, resulting in a step backwards, and also allowed the OpenAI team to successfully complete the double kill. Then, the AI team drummed up and successfully won a tower in the middle of the road. This wave gives the dwarf helicopter a good chance to develop, directly "killing and killing"!
The second wave of wonderful GANK, the tide of four people, the human player's beautiful return to the carbine, curbed the pace of AI forward, the situation held, OpenAI this wave of blood loss! (Attention!) OpenAI has learned to type! ）
The third wave of wonderful GANK, this time, the human player finally can't stand it and is ready to fight back!
The human player took the lead and zoomed in to control three people, but because the AI response was really a bit fast, the counterattack battle did not get a particularly large advantage.
In the fourth wave of wonderful GANK, the AI teams wandered each other, the human teams began to fight in groups, and guerrilla tactics annihilated the four members of OpenAI, almost annihilating the group. At this time, ai's strategy is to let the axe king single belt, the next road belt line to push the tower, but the effect is not very ideal.
In the later stage, OpenAI seems to have bugs, frequently interjecting at home, acting independently, and there is no cooperation between the teams at all, living like a headless fly. OpenAI is clearly unable to respond strategically in the face of headwinds, which once again proves that AI is far inferior to humans in terms of strategy formulation.
In the end, the human players won the game with a single blow!
Before the game, Greg Brockman, team leader of OpenAI Dota, said that according to the rule of three-to-two wins, if the AI loses this time, it will be their last game.
So, the Chinese players stopped the AI from breaking the path of dota.
Ai had pushed towers like mowing grass, but was unable to fight back in front of the average 9,000 points master
In the previous DOTA event, Open AI's performance was not so bad.
On August 5, five Dota 2 masters, who outrannumbered 99.5% of the world's players, played against OpenAI Five.
As a result, humans are abused by AI.
The final score was 2:1, and the humans won a round to save face.
But this sentence to save face is also due to the temporary rise of the organizers - perhaps three times in a row to watch the AI crush the human race for a short time, which is not so interesting to the live audience and the 100,000+ netizens watching the live broadcast.
In the third game, the audience assigned the AI a hero – and it was a bad combination, and the AI predicted that after determining the hero, the probability of winning was only 37%.
Yesterday, OpenAI Five debuted at TI8, the best tournament in the Dota 2 international major event, against the world's top best pro players (average 9,000 points), but the AI lost.
Shortly after yesterday's game began, OpenAI Five has been choosing to kill people instead of pushing the tower, Chinese Douyu commentary also ridiculed: Looking at it this way, I can also go to OpenAI and brag about X.
Rapidly evolving Open AI: Represents the highest international level of multi-agent decision-making intelligence
Since beating the world's top player Dendi in 1v1 at TI7 in 2016, OpenAI's Battle of Dota 2 Raiders has progressed rapidly: the evidence is the example above.
Just one year later, it can be successfully upgraded from 1v1 to 5v5, and Yuan Quan, founder and CEO of Qiyuan World, the first decision-making intelligence company in China, commented:
"The Dota game is a typical AI puzzle, which combines a long decision-making cycle, a large space, and the enemy and us are playing under incomplete information.
After solving the problem of 1v1 last year, OpenAI can show a long-term planning and collaboration ability similar to that of people under the premise of relying entirely on self-confrontation learning and no explicit communication channel in the more complex situation of 5v5 within 1 year, representing the highest international level of multi-agent decision-making intelligence, and also reflecting the beauty brought by large-scale computing power. ”
OpenAI Five also lived up to expectations, from a team with an average of 4,000 points at the ladder level to a human player who scored an average of 7,000 points, and it only took a few weeks of training time.
OpenAI Five scored a team of human players with an average of 7,000 points
If AI wins today, it will be a huge victory and an important milestone in the history of the artificial intelligence and gaming industry.
Before the game, the TI8 forum also launched a vote to let netizens pre-open AI win rate, unexpectedly, most people still chose to win human - and according to OpenAI's performance along the way, it seems that winning on TI8 is a breeze for AI.
Then, the situation was completely reversed — this time, it was the AI that was swept away by the human team.
Beat OpenAI Five's professional team pg
It all felt a bit abrupt.
AI has the gameplay of AI, and one day of training is equivalent to 180 years of human beings
OpenAI Five can't play full Dota 2 yet, and when it beats a team of prospective pros on August 5th, OpenAI Five can only choose 18 of the 115 heroes to play.
But OpenAI Five still exhibits precise calculations, aggressive fighting styles, and unstoppable momentum. The AI is amazing, not following the conventions of Dota 2, using new methods to control heroes, allocate resources, and ultimately win.
One reason AI plays so differently from humans is that AI is not human: AI can perform calculations far beyond top pro players, achieving superhuman efficiency and precision. What's even more appealing, however, is the cooperation between the 5 AI, openAI Five does not use rules to code, but trains itself through reinforcement learning.
Every second, OpenAI Five receives more than 20,000 observations from the Dota 2 API. These numbers describe everything from the hero's health to the number of seconds consumed by a particular enemy's debuff. At the same time, the 5 AIs can choose from thousands of actions – eye-catching, moving, attacking, using spells or items that appear on the map to various targets or locations.
The challenge for AI is to figure out which of these 20,000 observations is important right now, and which actions to take accordingly are more likely to help them win.
OpenAI Five predicts action in advance
The key to this is the neural network that brings all these inputs and outputs together. One of the most important functions of these neural networks is to give weight to each input —to increase or decrease the effect of a particular input on a particular output.
After losing the first game to the human professional team pg, OpenAI researchers said there were 3 main reasons for the loss, including the setting of the weights. In the next battle, they will adjust this.
Yesterday's game openai five economic situation
OpenAI Five starts with randomly selected weights, so anyone who has never played Dota 2 at this time has a chance to beat Five at this stage. But over time, the AI receives feedback — rewards when they get coins or kill enemy heroes, and punishment when their own heroes die — and that they adjust their weight slightly on their neural networks each time. Soon, the AI's ability to play Dota 2 will surpass that of ordinary people, after all, one day of training is equivalent to 180 years of human beings.
In the first game of ti8 played yesterday, the OpenAI Five performed well in terms of teamwork and mutual support, which can be seen in terms of the number of heads, and then worse on the economic side, initially stumbling and then completely collapsing.
OpenAI Five's predictions of their own wins at the start of the battle have been high, and they are determined to play aggressive game styles, until the AI still exceeds 70% of their chances of winning until they finally lose.
It also makes it seem understandable that the AI's series of actions are more likely to win, and I can do more interjection and explore various places on the map. Planning long-duration games requires anticipating the longer-term future and being able to separate activities for 10, 20 or 30 minutes.
It turns out that collecting enough gold in 30 minutes to become powerful enough to win a game is harder to learn than unleashing a skill to kill a hero.
The end of the game AI is also a new beginning
In the second game, OpenAI Five went through some tweaks to fix the so-called "bug" that was shown in one game...,
For AI researchers, victory is victory, which will be another milestone in the history of artificial intelligence.
But on the other hand, for Dota2 or the esports gaming industry as a whole, the impact is small at the moment — OpenAI's approach takes months of training, as well as millions of dollars in hardware, not to mention a bunch of the world's top AI researchers.
For OpenAI, beating humans in Dota 2 is the first step in the long journey to universal AI in the real world.
Where is that to celebrate?
What's the use of OpenAI getting better on Dota 2, or DeepMind suddenly developing the best StarCraft 2 AI in the world?
First, it will remind us that, like AlphaGo inspired the Go community back then, the games humans play every day still contain unknown dimensions.
OpenAI Five has superhuman reactions, but the heroes cooperate with each other and the tactic of sending four heroes to push the tower in the first minute. AI will always challenge our understanding, keep humans looking for new strategies, and provide us with a goal that we can continue to work towards.
The victory (or defeat) of OpenAI Five does not represent the end of the game AI, but a new beginning.
Popular Articles:
The new Zhiyuan AI WORLD 2018 conference countdown is 26 days
Tickets are on sale!
Xinzhiyuan will hold AI WORLD 2018 conference at the Beijing National Convention Center on September 20, inviting the godfather of machine learning and CMU professors Tom Mitchell, Max Tagmark, Zhou Zhihua, Tao Dacheng, Chen Yiran and other AI leaders to focus on machine intelligence and human destiny.
Official website:
http://www.aiworld2018.com/
Link to purchase tickets for the event line:
http://www.huodongxing.com/event/6449053775000
QR code for ticket purchase: