DeepMind sets AlphaGo's sights on SCII - Page 6

On March 29 2016 02:19 HellHound wrote:

EPM not APM.

On March 28 2016 12:53 Circumstance wrote:
The real-time aspect will be critical. Go, being turn-baseed, is purely a matchup of mind against "mind". If They allow the machine unlimited APM, then this won't be much of a match.

Also, you'd need representatives from all 3 races, wouldn't you?

On March 29 2016 03:22 Incognoto wrote:


Yeah I think it's possible that deepmind wins solely off of micro.

Unless they implement an APM limiter or something

On March 29 2016 02:14 Mendelfist wrote:

Why do you assume that the best way to solve this problem is to throw every single game state variable or pixel at a neural net and then hope that it somehow works out?



Then imagine one. One that learns for example by self play. Do you really think that it would have a hard time finding "the right moves" just because they are infinite in number? Edit: And the question is not if it can be better than a script, which can be near perfect. The question is if you think it's a hard problem.


Yes, THIS is the problem, and once you have done this the "number of possible moves" in the original problem is irrelevant. That only tells you that you have THIS problem on your hands and that you can't solve it by ordinary search algorithms. You have to find a way to reduce the original problem by levels of abstraction. I don't know if there is a way for an AI find these abstractions by itself. Maybe that's how Alphago works, In any case, I'm not in the least convinced that it's as hard as you are trying to make it sound. In the simplest form you can have an ordinary scripted bot that asks an AI for advice. "Attack now", "build sentries", "expand there" etc. Or you could throw every pixel at it, like you want. I don't think that would work. Or something in between. How about that?

On March 29 2016 03:49 The Bottle wrote:
The more complex the game is (and I mean this in terms of permutations of moves, not in terms of heuristic strategy), the harder this task is. Thus it will be monumentally difficult for Starcraft.

it is incredibly difficult to reduce the space in an intelligent enough way to minimally reduce the information of best moves possible

On March 29 2016 02:39 diabcockiful wrote:



Yeah, but SC2 requires a little more creativity than GO or Chess...I think eventually an AI could be better of course, but I don't think we're anywhere near that point yet. There aren't predetermined tiles or spots for pieces to move...SC2 units don't have simple rules, and there are almost infinite possibilities. I suppose you could teach an AI to macro like an animal and then A-move over someone, or maybe even attempt to setup concaves. But man, there are a lot more factors involved in winning an SC2 engagement than taking someone's rook in chess. And while I don't know the rules of Euro.Go, it looks like you have one type of unit and a grid to move on... No setting up army compositions, or gaining vision of your opponent, etc.

Maybe I'll be surprised as to how far AIs have come, but methinks they will be mechanically perfect but fairly retarded in many ways.

On March 29 2016 03:49 The Bottle wrote:
But listen, because I think you're still misunderstanding me. I know that they're taking shortcuts to greatly reduce the search space of possible moves in Go, the paper states this pretty clearly. But the problem is still that, because of the sheer number of possible moves you can make, and the stark difference in outcome between those moves, it is incredibly difficult to reduce the space in an intelligent enough way to minimally reduce the information of best moves possible. The more complex the game is (and I mean this in terms of permutations of moves, not in terms of heuristic strategy), the harder this task is. Thus it will be monumentally difficult for Starcraft. Yes, they will have to take new shortcuts, of the sort that they didn't take in Go. But every time they do such a thing, they have to be incredibly careful not to remove certain specific crucial moves, or coarse grain it in with other moves that have drastically different results. (An example of this is in zvz ling bane wars, where a couple pixels different motion can be the difference between 2 dead lings and 20 dead lings, or aiming a disruptor shot, or things like that).

On March 29 2016 04:48 kingjames01 wrote:
Beyond a ladder (a large board-scale trap) calculator, the DeepMind team did not provide AlphaGo any hints, tips or tricks. Through many iterations, AlphaGo learned to place stronger weights on specific paths to search.

On March 29 2016 04:44 Musicus wrote:


Since they are already talking, I'm sure Blizzard will provide Google with whatever they need. The publicity for sc2 will be insane.

On March 29 2016 05:00 Mendelfist wrote:

I'm going off a slight tangent here, but someone on the DeepMind team spoke of a possible future development, and that would be doing Alphago again but without the first training step of the neural nets with lots of human games. While they didn't teach it any specific tricks, these games may have taught it bad habits. Go is actually a very little researched game, which Go Seigen proved in the middle of the last century by turning everything upside down. I would very VERY much want them to do this instead of trying Starcraft. It could have vast implications for our understanding of Go. Maybe the best starting move is right in the middle?