| Forum Index > SC2 General |
|
MyLovelyLurker
France756 Posts
On November 05 2016 08:38 jimminy_kriket wrote: wouldnt that be 180 actions per second? That would be cool to see. They stacked frames in Atari to have the algorithm see animations, I'll assume they'll do the same here. 180 actions per second is 10,000 APM, that'd be too much. | ||
|
Tuczniak
1561 Posts
On November 05 2016 08:38 jimminy_kriket wrote: Yes. But it shouldn't be hard to limit it to any reasonable number.wouldnt that be 180 actions per second? That would be cool to see. | ||
|
LetaBot
Netherlands557 Posts
An example of these algorithms in action: | ||
|
imp42
398 Posts
Who said you'd have to evaluate every single possible move to become good at the game? It's obviously not what we humans do. In fact, this is the message I attempted to get across in my latest post, where I took the idea to the other extreme, reducing the game state to 6 variables and the number of possible moves to 4: http://www.teamliquid.net/blogs/515461-towards-a-good-sc-bot-p4-planning-1-3 as it usually does, the sweet spot lies somewhere in between the two extremes... | ||
|
Jett.Jack.Alvir
Canada2250 Posts
| ||
|
Cyro
United Kingdom20326 Posts
On November 05 2016 07:42 MyLovelyLurker wrote: If Deepmind proceeds as they have in the past for Atari games AI, their SC2 AI would automatically be limited to 2 or 3 actions per frame, so at 60fps, that's max 180APM. The sc2 simulation runs at ~22hz regardless of framerate so that may be a relevant number | ||
|
LetaBot
Netherlands557 Posts
https://gist.github.com/spencerduncan/c1d383819e032ab61cb266d8d283a67f | ||
|
JWD[9]
364 Posts
With the Atari games I thought they just used keyboard input and pixelvalues of the screen, why not do the same for sc2? I don't know, but do people that are familiar with the BW API know if opening up that SC2 API will make hacks more prevelant? | ||
|
Dumbledore
Sweden725 Posts
On November 06 2016 05:23 JWD[9] wrote: If they need the API to program the interface for the AI, then the AI is cheating. Human brains have to work through visual clutter aswell. Will the Ai have to zoom in to snipe units under collosi? Will it have to rotate the camera to check for widowmines behind extractors? With the Atari games I thought they just used keyboard input and pixelvalues of the screen, why not do the same for sc2? I don't know, but do people that are familiar with the BW API know if opening up that SC2 API will make hacks more prevelant? Lol I've never heard of anyone zooming in to snipe units under a colossus. | ||
|
LetaBot
Netherlands557 Posts
On November 06 2016 05:23 JWD[9] wrote: If they need the API to program the interface for the AI, then the AI is cheating. Human brains have to work through visual clutter aswell. Will the Ai have to zoom in to snipe units under collosi? Will it have to rotate the camera to check for widowmines behind extractors? With the Atari games I thought they just used keyboard input and pixelvalues of the screen, why not do the same for sc2? I don't know, but do people that are familiar with the BW API know if opening up that SC2 API will make hacks more prevelant? As mentioned in the talk at Blizzcon, there will be an API that uses visual information and an API that works like the BWAPI for Brood War The SC2 API won't be capable to be used for ladder games. | ||
|
paralleluniverse
4065 Posts
| ||
|
JWD[9]
364 Posts
On November 06 2016 06:08 LetaBot wrote: As mentioned in the talk at Blizzcon, there will be an API that uses visual information and an API that works like the BWAPI for Brood War The SC2 API won't be capable to be used for ladder games. Ah, then I am on board, go ahead, ForScience (and more sc2 longlivity  ) | ||
|
sabas123
Netherlands3122 Posts
On November 06 2016 06:49 paralleluniverse wrote: SC2 can't be solved. I'd expect the AI to find some near unbeatable early game cheese, and ultimately it'll probably reduce to randomizing between executing a few strategies perfectly. Games don't have to be solvable for an AI to play it better than any person. GO is an perfect example of that. | ||
|
BluzMan
Russian Federation4235 Posts
On March 28 2016 13:39 JimmyJRaynor wrote: without heuristic functions i think AlphaStar will have trouble "learning" the game. and with heuristic functions it'll play with the style of dictated by them. The whole point of unsupervised learning is the AI picking up the right heuristics by itself. In the match against Lee Se Dol, AlphaGo made an attack that no pro could predict in a game that it won. It saw something in the board state that made it do the winning move, because some heuristic it developed "triggered" on said board state. EDIT: I actually think that the best advantage of a potential AI is individual unit control and AoE-immune spreads. StarCraft 2 default AI (the one that players trigger by rightclicking) is VERY BAD in aoe-related pathfinding, so that alone might be enough to make all other advantages irrelevant. | ||
|
Xamo
Spain886 Posts
| ||
|
Lexender
Mexico2656 Posts
On November 06 2016 08:03 BluzMan wrote: EDIT: I actually think that the best advantage of a potential AI is individual unit control and AoE-immune spreads. StarCraft 2 default AI (the one that players trigger by rightclicking) is VERY BAD in aoe-related pathfinding, so that alone might be enough to make all other advantages irrelevant. They are not going to do that tho, they aren't going to spend a ton of money in a super AI only to make it be a automaton 1000 bot that only makes perfectly microed marines. | ||
|
mostevil
United Kingdom611 Posts
| ||
|
Hider
Denmark9433 Posts
This restriction will mean that they need to beat pro's through strategy. | ||
|
Proko
United States1022 Posts
| ||
|
lolfail9001
Russian Federation40190 Posts
| ||
| ||
|
|
Replay Cast
GSL
Cure vs TriGGeR
ByuN vs Bunny
KCM Race Survival
Big Gabe
Replay Cast
Replay Cast
Escore
OSC
Replay Cast
Replay Cast
[ Show More ] |
|
|
EPT
