OpenAI's Dota 2 agents display their dominance over mere mortals one last time

[Cal Jeffrey]

Bottom line: After handily defeating Dota 2 pros last year, OpenAI decided to allow amateur gamers to face off against its Dota playing bots (called ‘agents’) before retiring the team. The results were predictably lopsided.

In total, the AI team, called OpenAI Five, played 7,257 matches and only lost 42 for a win ratio of 99.4 percent. The agents won 4,075 games outright. Human players quit, admitting defeat, 3,140 times. OpenAI Five’s dominance against amateur players is not surprising.

Last summer the bots were pitted against a team of professional Dota players and defeated them 2-1 in a best of three series. Even though the AI did lose one match, it was constrained to a choice of only 18 of the 100 heroes and had its reaction time increased to 200ms from its usual 80ms.

More recently, OpenAI’s agents were tested not only against professionals but with The International 2018 champion, team OG. In the best-of-three match, OpenAI Five swept OG in the first two games making a third round unnecessary.

The Verge notes that at the time of the pairing between OpenAI Five and OG, the agents had already had the equivalent of 45,000 years worth of training.

“OpenAI Five is powered by deep reinforce learning, which means we didn’t code it how to play. We coded it how to learn,” said Greg Brockman, OpenAI’s CTO. “In its 10 months of existence, it’s already played 45,000 years of Dota 2 gameplay. That’s a lot — it hasn’t grown bored yet.”

The OpenAI bots can also play in cooperative mode with humans. Logically, the amateur players enjoyed playing with the agents, more than playing against them. The Verge notes that the cooperative version of the AI received far more attention with more than 35,000 games played.

For now, OpenAI Five is retired. The team will review the data accumulated from the matches and “will use Dota 2 as a research platform for further development of powerful AI systems.”

Permalink to story.

https://www.techspot.com/news/79768-openai-dota-2-agents-display-their-dominance-over.html

 

[ShagnWagn]

So the whole idea behind "AI" (which was created from human thought - not 'artificial') is not how fast it reacts, but intelligence. Letting it react as fast as a computer is not a rating of intelligence. This is like pitting a human against a robot in strength. This was very obvious when they first brought openai against people. Real human intelligence won over sheer power.

"had its reaction time increased to 200ms from its usual 80ms."

Is 200ms typical of real players? If so, then it would really be a fair contest of human-programmed intelligence to real-time human intelligence.

If "AI" really is intelligent, why are phones so stupid? When will we have AI phones instead of "smart" ID10T phones.

 

[syrious01]

So the whole idea behind "AI" (which was created from human thought - not 'artificial') is not how fast it reacts, but intelligence. Letting it react as fast as a computer is not a rating of intelligence. This is like pitting a human against a robot in strength. This was very obvious when they first brought openai against people. Real human intelligence won over sheer power.

"had its reaction time increased to 200ms from its usual 80ms."

Is 200ms typical of real players? If so, then it would really be a fair contest of human-programmed intelligence to real-time human intelligence.

If "AI" really is intelligent, why are phones so stupid? When will we have AI phones instead of "smart" ID10T phones.

I guess if you do some extremely dirty math you can see where AI has the learning advantage.
Average game = 1hr
Hours in a year = 8760hrs (not factoring in sleeping, pissing, and eating)
Potential AI games in 45,000 years = 394,200,000

Even if you played dota every minute of every date until you were 35, you'd still only have 306,600 hours of experience.

 

[ShagnWagn]

I guess if you do some extremely dirty math you can see where AI has the learning advantage.
Average game = 1hr
Hours in a year = 8760hrs (not factoring in sleeping, pissing, and eating)
Potential AI games in 45,000 years = 394,200,000

Even if you played dota every minute of every date until you were 35, you'd still only have 306,600 hours of experience.

I guess I did not explain the overall point very well.

Yours: 394,200,000 games and it still is getting beaten? Ouch, this explains one of my points.

 

[J spot]

So the whole idea behind "AI" (which was created from human thought - not 'artificial') is not how fast it reacts, but intelligence. Letting it react as fast as a computer is not a rating of intelligence. This is like pitting a human against a robot in strength. This was very obvious when they first brought openai against people. Real human intelligence won over sheer power.

"had its reaction time increased to 200ms from its usual 80ms."

Is 200ms typical of real players? If so, then it would really be a fair contest of human-programmed intelligence to real-time human intelligence.

If "AI" really is intelligent, why are phones so stupid? When will we have AI phones instead of "smart" ID10T phones.

I don't think you truly understand the concept of AI, yet.

 

[gamerk2]

So the whole idea behind "AI" (which was created from human thought - not 'artificial') is not how fast it reacts, but intelligence. Letting it react as fast as a computer is not a rating of intelligence. This is like pitting a human against a robot in strength. This was very obvious when they first brought openai against people. Real human intelligence won over sheer power.

"had its reaction time increased to 200ms from its usual 80ms."

Is 200ms typical of real players? If so, then it would really be a fair contest of human-programmed intelligence to real-time human intelligence.

If "AI" really is intelligent, why are phones so stupid? When will we have AI phones instead of "smart" ID10T phones.

The problem with current AI is that no implementation is really learning. The current preferred method is reinforcement learning, which does an OK job of making a competent AI that can do one specific thing, but also has significant limitations (bad behaviors that are learned early will never be unlearned) and can be exploited once understood.

For example, the SC2 AI lost a game because the human player figured out the AI would disengage it's entire army to defend it's base, even if just one unit was attacking it. This was a learned behavior, and the human punished it by preventing the AI from attacking until the human could win through sheer numbers. A "smart" AI would have figured this out, but the AI was bound by it's past experiences and didn't have any "thought" to see, in realtime, what was happening.

Older AIs simply used databases they would access to determine what the proper results should be. Most chess engines fall into this territory, as does Siri/Alexia and the most other commercial implementations. They do OK at what they do, but a very limited in function as a result.

An AI with actual independent thought is well beyond what we are capable of at the moment. Putting aside the processing and storage requirements, no one is even in agreement on the best way to do that. That is honestly just as much a sociological question as an engineering one.

 

[shm0wie]

So the whole idea behind "AI" (which was created from human thought - not 'artificial') is not how fast it reacts, but intelligence. Letting it react as fast as a computer is not a rating of intelligence. This is like pitting a human against a robot in strength. This was very obvious when they first brought openai against people. Real human intelligence won over sheer power.

"had its reaction time increased to 200ms from its usual 80ms."

Is 200ms typical of real players? If so, then it would really be a fair contest of human-programmed intelligence to real-time human intelligence.

If "AI" really is intelligent, why are phones so stupid? When will we have AI phones instead of "smart" ID10T phones.

Comparing AI to the software in phones is like comparing a model airplane to an F18, and asking why the model airplane won't fly.

 

[J spot]

The problem with current AI is that no implementation is really learning. The current preferred method is reinforcement learning, which does an OK job of making a competent AI that can do one specific thing, but also has significant limitations (bad behaviors that are learned early will never be unlearned) and can be exploited once understood.

For example, the SC2 AI lost a game because the human player figured out the AI would disengage it's entire army to defend it's base, even if just one unit was attacking it. This was a learned behavior, and the human punished it by preventing the AI from attacking until the human could win through sheer numbers. A "smart" AI would have figured this out, but the AI was bound by it's past experiences and didn't have any "thought" to see, in realtime, what was happening.

Older AIs simply used databases they would access to determine what the proper results should be. Most chess engines fall into this territory, as does Siri/Alexia and the most other commercial implementations. They do OK at what they do, but a very limited in function as a result.

An AI with actual independent thought is well beyond what we are capable of at the moment. Putting aside the processing and storage requirements, no one is even in agreement on the best way to do that. That is honestly just as much a sociological question as an engineering one.

There are two top AI chess engines, Alpha Zero and Leela Zero, and they are much more impressive than traditional brute force chess engines. Their "thinking" is phenomenal. And as far as a brute force chess engines, like Stockfish, no human that ever lived has a chance against it.

 

[gamerk2]

There are two top AI chess engines, Alpha Zero and Leela Zero, and they are much more impressive than traditional brute force chess engines. Their "thinking" is phenomenal. And as far as a brute force chess engines, like Stockfish, no human that ever lived has a chance against it.

All those engines do is essentially play themselves over and over, acquiring a massive database of past knowledge that is used to make the correct move in any situation. As I noted, this is the way AIs are currently moving toward, and even then mistakes are made from time to time (and yes, I watched the entire Alpha Zero/Stockfish series; I noted some moves from both engines that appear sub-optimal).

 

[J spot]

All those engines do is essentially play themselves over and over, acquiring a massive database of past knowledge that is used to make the correct move in any situation. As I noted, this is the way AIs are currently moving toward, and even then mistakes are made from time to time (and yes, I watched the entire Alpha Zero/Stockfish series; I noted some moves from both engines that appear sub-optimal).

I can't take the comment that you saw some moves from both engines that appeared sub-optimal. There's just no way for you as a human, who's not even a professional that dedicates their whole life to chess and chess theory, and plays eight hour games of pure thinking and evaluating, is in any position to know whether any move that Alpha Zero made was sub optimal or not.