It both is and isn't. It appears that though the AI generalization aspect of the project is coming along nicely (winning games with less than perfect knowledge) it still has a way to go when it's...
It both is and isn't.
It appears that though the AI generalization aspect of the project is coming along nicely (winning games with less than perfect knowledge) it still has a way to go when it's knowledge is more limited like a human's. The last game was really the thing that solidified this view as it was able to only see one part of the map at a time but also not have complete knowledge of unit status that it wasn't able to see.
Though I don't see anything particularly nefarious with setting up the games this way initially it definitely makes a strong case that this was set up in such a way to maximize the computer's competiveness while still being "fair". I don't really take issue with the fact with the level of perfection that is programmed into it's unit selection and other aspects of "manual" control since humans can and do perform at levels as close to perfection as possible in a lot of aspects already (git gud).
Overall though very encouraging to see, definitely not a Deep Blue-Kasparov moment but still very exciting to see.
I watched the stream and honestly it did feel like the Alpha star team did a lot of 'meta gaming' to eek out an advantage. For instance both players had no idea what they would be facing, then...
I watched the stream and honestly it did feel like the Alpha star team did a lot of 'meta gaming' to eek out an advantage. For instance both players had no idea what they would be facing, then they did not tell them about the agent system so they wasted a match changing strategies when that is completely irrelevant as the new agent would never play the same as the one before.
Oh most definitely, this was a high stakes public demonstration of the capabilities of Alpha Star anything short of over 50% win rate would have really rained on their parade. They had to make...
Oh most definitely, this was a high stakes public demonstration of the capabilities of Alpha Star anything short of over 50% win rate would have really rained on their parade. They had to make sure that they minimized the disadvantages of AlphaStar while maximizing its advantages.
That being said it was still a eye opening demonstration of how far this technique has progressed.
It's particulary interesting to read the wall of text below the video. Their description how the AI discovered new strategies, and abandoned previous ones - just like human progression - really...
It's particulary interesting to read the wall of text below the video.
Their description how the AI discovered new strategies, and abandoned previous ones - just like human progression - really sounds ... weird. Their comparsion of AI to humans feels really weird, like if it was beginning of something new, and altough I know this is, while great achievement, "just* a big improvement of previous AIs, the way they said really made me feel weird. This is the first time I really consider an AI smart.
“I was impressed to see AlphaStar pull off advanced moves and different strategies across almost every game, using a very human style of gameplay I wouldn’t have expected,” he said. “I’ve realised how much my gameplay relies on forcing mistakes and being able to exploit human reactions, so this has put the game in a whole new light for me. We’re all excited to see what comes next.”
“I was surprised by how strong the agent was,” he said. “AlphaStar takes well-known strategies and turns them on their head. The agent demonstrated strategies I hadn’t thought of before, which means there may still be new ways of playing the game that we haven’t fully explored yet.”
Part of this is most likely due to the agents being trained on human data, it seems like the objective (to an extent) was in fact to make them play in a way which resembled human players.
Part of this is most likely due to the agents being trained on human data, it seems like the objective (to an extent) was in fact to make them play in a way which resembled human players.
This is not true in practice, if only because you can actually create a "starting point" for the bots by simulating human-like data using statistical inferences from the human data. But more...
This is not true in practice, if only because you can actually create a "starting point" for the bots by simulating human-like data using statistical inferences from the human data. But more pertinently because you can train networks in adversarial setups, and they can very much learn to perform up to spec or even better than humans in such a case, without ever involving human data.
Deepmind has also been applying "AlphaZero" to other games recently, where it learns to play the game completely from scratch with no training data at all:...
I don't know if it's feasible to apply the same approach to StarCraft II, but it enabled it to learn to play chess/go/shogi without any training data except what it produced itself.
I think the big limit on the AlphaZero approach to Starcraft is simply that they first needed to establish whether or not a network could reliably identify features in the data and what those...
I think the big limit on the AlphaZero approach to Starcraft is simply that they first needed to establish whether or not a network could reliably identify features in the data and what those features were, it stands to reason that they will at least attempt a similar approach now that they have some idea of what a 'Zero' agent needs to operate in Starcraft 2, but there's also a question whether or not that's currently useful or they wish to move on to other projects.
I'd be interested if the AI's limitations were made to be closer to a human in a physical sense, rather than a statistical sense. What I mean is that a human has two arms that must move both...
I'd be interested if the AI's limitations were made to be closer to a human in a physical sense, rather than a statistical sense. What I mean is that a human has two arms that must move both quickly and accurately.
Some strategies of the AI seem to require super-human accuracy in clicking, which isn't reflected in the APM counter. The example is microing the damaged units to flash to the back of the group. It seems like this would require mouse accuracy that is beyond what a human could do repeatedly. Compare this to keyboard hotkeys where a human can pretty reliably press the required key.
I'd be curious is the AI uses less keyboard shortcuts than a human because the mouse accuracy is so high that there is no penalty to using the mouse.
This is an interesting article that contends that they can't make it more human-like: An Analysis On How Deepmind’s Starcraft 2 AI’s Superhuman Speed is Probably a Band-Aid Fix For The Limitations...
Interesting read. I might have to look at some starcraft communities to see what their thoughts are. One problem I've seen so far is people that don't know much about AI are humanising the AI too...
Interesting read. I might have to look at some starcraft communities to see what their thoughts are.
One problem I've seen so far is people that don't know much about AI are humanising the AI too much, assuming that it 'thinks and feels' and learns like a human, where in reality it is based purely on statistical information (win/loss), which funnily enough tends to be one of the weakest traits in humans. The article you linked seems to be quite fluent in that area, which is unfortunately quite rare in this topic.
This is a pretty big deal, right?
I'm surprised I didn't see anything about it when the matches took place in December.
It both is and isn't.
It appears that though the AI generalization aspect of the project is coming along nicely (winning games with less than perfect knowledge) it still has a way to go when it's knowledge is more limited like a human's. The last game was really the thing that solidified this view as it was able to only see one part of the map at a time but also not have complete knowledge of unit status that it wasn't able to see.
Though I don't see anything particularly nefarious with setting up the games this way initially it definitely makes a strong case that this was set up in such a way to maximize the computer's competiveness while still being "fair". I don't really take issue with the fact with the level of perfection that is programmed into it's unit selection and other aspects of "manual" control since humans can and do perform at levels as close to perfection as possible in a lot of aspects already (git gud).
Overall though very encouraging to see, definitely not a Deep Blue-Kasparov moment but still very exciting to see.
I watched the stream and honestly it did feel like the Alpha star team did a lot of 'meta gaming' to eek out an advantage. For instance both players had no idea what they would be facing, then they did not tell them about the agent system so they wasted a match changing strategies when that is completely irrelevant as the new agent would never play the same as the one before.
Oh most definitely, this was a high stakes public demonstration of the capabilities of Alpha Star anything short of over 50% win rate would have really rained on their parade. They had to make sure that they minimized the disadvantages of AlphaStar while maximizing its advantages.
That being said it was still a eye opening demonstration of how far this technique has progressed.
I think some of the "shine" of it wore off after the Dota 2 project, since this reads like more of the same to most people (I assume).
It's particulary interesting to read the wall of text below the video.
Their description how the AI discovered new strategies, and abandoned previous ones - just like human progression - really sounds ... weird. Their comparsion of AI to humans feels really weird, like if it was beginning of something new, and altough I know this is, while great achievement, "just* a big improvement of previous AIs, the way they said really made me feel weird. This is the first time I really consider an AI smart.
Part of this is most likely due to the agents being trained on human data, it seems like the objective (to an extent) was in fact to make them play in a way which resembled human players.
Well, it's not like you could train them on nonhuman data since the other bots all suck.
This is not true in practice, if only because you can actually create a "starting point" for the bots by simulating human-like data using statistical inferences from the human data. But more pertinently because you can train networks in adversarial setups, and they can very much learn to perform up to spec or even better than humans in such a case, without ever involving human data.
Deepmind has also been applying "AlphaZero" to other games recently, where it learns to play the game completely from scratch with no training data at all: https://www.newyorker.com/science/elements/how-the-artificial-intelligence-program-alphazero-mastered-its-games
I don't know if it's feasible to apply the same approach to StarCraft II, but it enabled it to learn to play chess/go/shogi without any training data except what it produced itself.
I think the big limit on the AlphaZero approach to Starcraft is simply that they first needed to establish whether or not a network could reliably identify features in the data and what those features were, it stands to reason that they will at least attempt a similar approach now that they have some idea of what a 'Zero' agent needs to operate in Starcraft 2, but there's also a question whether or not that's currently useful or they wish to move on to other projects.
I'd be interested if the AI's limitations were made to be closer to a human in a physical sense, rather than a statistical sense. What I mean is that a human has two arms that must move both quickly and accurately.
Some strategies of the AI seem to require super-human accuracy in clicking, which isn't reflected in the APM counter. The example is microing the damaged units to flash to the back of the group. It seems like this would require mouse accuracy that is beyond what a human could do repeatedly. Compare this to keyboard hotkeys where a human can pretty reliably press the required key.
I'd be curious is the AI uses less keyboard shortcuts than a human because the mouse accuracy is so high that there is no penalty to using the mouse.
This is an interesting article that contends that they can't make it more human-like: An Analysis On How Deepmind’s Starcraft 2 AI’s Superhuman Speed is Probably a Band-Aid Fix For The Limitations of Imitation Learning
Interesting read. I might have to look at some starcraft communities to see what their thoughts are.
One problem I've seen so far is people that don't know much about AI are humanising the AI too much, assuming that it 'thinks and feels' and learns like a human, where in reality it is based purely on statistical information (win/loss), which funnily enough tends to be one of the weakest traits in humans. The article you linked seems to be quite fluent in that area, which is unfortunately quite rare in this topic.