probably a (big) incremental step over a previous breakthrough
Is AlphaZero really a breakthrough in AI?
31–40 of 84 posts
Re: Is AlphaZero really a breakthrough in AI?
#32I'm no AI expert, but I won't start to worry about AI being used generally (the last point of the article) until it beats a really complex game like real time strategy title StarCraft.
Even if we all agree that AlphaGo is the DeepBlue of Go, we are still having a few more layers to take before humans need to worry.
Re: Is AlphaZero really a breakthrough in AI?
#33I'm no AI expert, but I won't start to worry about AI being used generally (the last point of the article) until it beats a really complex game like real time strategy title StarCraft. Even if we all agree that AlphaGo is the DeepBlue of Go, we are still having a few more layers to take before humans need to worry.
> Churchill guesses it will be five years before a StarCraft bot can beat a human. He also notes that many experts predicted a similar timeframe for Go—right before AlphaGo burst onto the scene.
https://www.wired.com/story/googles-ai-declares-galactic-war...
(Aug 2017)
Re: Is AlphaZero really a breakthrough in AI?
#34Earlier quoted context omitted.
If possible at all, could you give an example for casual chess players?
Seconded. We all want a glimpse of what the "AI" is doing, and see that one step ahead that human minds could never do :)
Not OP, but I highly recommend checking out ChessNetwork's recent videos on the games: https://www.youtube.com/user/ChessNetwork/videos
Re: Is AlphaZero really a breakthrough in AI?
#35Whether it was a breakthrough or not, I have to say, the moves it played were certainly "creative" in a profound sense.
If possible at all, could you give an example for casual chess players?
I would recommend some analysis videos by
agadmator. Here's a good one: https://www.youtube.com/watch?v=lb3_eRNoH_w
Re: Is AlphaZero really a breakthrough in AI?
#36Whether it was a breakthrough or not, I have to say, the moves it played were certainly "creative" in a profound sense.
If possible at all, could you give an example for casual chess players?
International Master Daniel Rensch made a very informative YouTube series about the match, I highly recommend it. Here are the highlights:
Re: Is AlphaZero really a breakthrough in AI?
#37The article glosses over why the 4 hours was possible. Firstly, a major challenge in training an AI of this sort is getting enough labelled data. They played 300,000 games from memory. Under normal circumstances, that requires access to 300,000 games played by experts so the AI can learn to copy what the export does. That is how Alpha Go did it. AlphaZero neatly side steps this by generating it's own training data by…
>>I don't know whether they customised the AI for each game, but I suspect if
the input and output layers were wide enough to accommodate the largest game
they could use the same one for each. The Markov engine does have to know how
to make all legal moves from any game position, but coding that isn't rocket
science or particularly time consuming. The AI does _not_ start out knowing
those rules - it learns them from the Markov engine. It's all very DRY.
Coding a set of rules for chess may not be a big deal. Encoding it into a neural network architecture, so that the network only explores valid game states, _is_ a big deal, like the article says.
The paper itself clarifies that the network architecture is matched to the game board grid. Each position and each move are represented by features described as sets of "stacked" planes (i.e. 2d feature vectors). So for example, in chess, the board is a set of 8 x 8 planes, and the movement of a piece is represented by a set of 56 planes for "queen moves" (continuous moves in any direction) and another 8 for "knight moves" (jumping over pawns). How they constraint specific pieces to subsets of those planes, is not explained, or I missed it.
That's not an obvious setup, neither is it a trivial one. More to the point, it's absolutely not something you can just copy and paste to an arbitrary other game- it can only be applied to a game that uses the same board and same pieces as chess. So, for example, no, you couldn't use that same architecture to represent a game of poker.
Re: Is AlphaZero really a breakthrough in AI?
#38Earlier quoted context omitted.
Self play reinforcement learning has been a thing for decades it really isn't new. That's how I wrote my Lines of Acton and Dominion AIs years ago. (Disclaimer: I had a sign flip bug in the Reinforcement Learning step of the Lines of Action bot that made it worse every generation! After a few thousand iterations it was _really_ bad at playing lines of action)
Lines of Acton sounds like an excellent game for Londoners. (Acton is a suburb of London and has more railway stations bearing its name than any other place in the UK)
Ah yes, a derivative of Mornington Crescent more suitable for computers to play.
Re: Is AlphaZero really a breakthrough in AI?
#39Can anyone summarise how self play works here if AlphaZero only starts out being told the rules of the game? Does it initially plays games using completely random moves as both players? Is it only told who the winner is with no other feedback? How is it able to learn e.g. that certain moves at the start eventually lead to a win?
Re: Is AlphaZero really a breakthrough in AI?
#40The article glosses over why the 4 hours was possible. Firstly, a major challenge in training an AI of this sort is getting enough labelled data. They played 300,000 games from memory. Under normal circumstances, that requires access to 300,000 games played by experts so the AI can learn to copy what the export does. That is how Alpha Go did it. AlphaZero neatly side steps this by generating it's own training data by…
>> Finally, he overstates how hard it is to customise the engine (Markov algorithm + AI) to a game.
I think you're imagining some setup in which a neural net is directly communicating with an external rule engine that generates a set of legal moves for each piece in each board configuration. That's not how it works. You know how people say that Neural Nets are "black boxes"? That's what they mean. You can't just plug into arbitrary steps of their training with arbitrary external processes. You have to setup everything right at the input and then wait patiently at the output, until what you want comes out. Then you can plug that in to whatever reasoning loop you want to set up.
So like I say, it's not a trivial thing and it doesn't generalise very well to arbitrary problems. You need to match the architecture to the problem. For example, that's why ConvNets have been studied extensively for vision- their architecture is specifically tailored to vision problems.
Hence also the doubts expressed in the article about the claims to general game-playing ability. Yeah, you might imagine that it's possible to create a specialised architecture to play any particular game (or even sets of games with similar setups, like chess and shogi)- but it's not clear how exactly you'd do that automatically. Handcrafting a thousand architectures for a thousand different games is not very "general purpose", is it?