AI boffins turn to StarCraft to train future neural networks StarCraft could be the next battleground for AI, as researchers create an open framework that tests deep-learning methods in the real-time strategy game. Teaching AI to play games is serious business. Games act like milestones; when a machine is superior to humans at playing difficult games, it’s a sign that its neural net has …
As long as the machines also have to operate meaty appendages that are using crude input devices to control the far flung units and buildings, then it might actually be a fair fight. If they get to plug straight into the game without any encumbrances then they're just poor sportsbots.
This has been done before. The AIs have the advantage of seeing the full map, whereas the player doesn't. Also many of the maps are not 100% equal for both sides. Initial positioning and first player to take a particular vantage point factors a lot into which bot wins, as does starting race. Certain maps with certain races combinations are a guaranteed win when both AIs are playing fully optimized. Also, even if playing identically on a 100% equal map, one of the AIs is always mathematically ahead of the other in what is essentially a turn based game. And as such one will always win out by miliseconds.
When did StarCraft become turn based? Not played that version before.
If it was 100% equal it does not mean the AIs would do equal things. If one got ahead of the other (assuming they are doing the exact same thing at the exact same time or very close to. What I assume you mean by turn based) then the AI that is behind would alter it's strategy would it not?
AI is not pre-programmed software that will only do A if B occurs. It is trying to look for the best scenario with the inputs it is given. If those inputs change (as they would if one got ahead) then so should the AI output.
No, because every matchup has so many variations, timing attacks, all ins, cheeses, macro, decisions about unit compositions, gas timings, when to build what, you need your scouting on point, deny scouting, early, mid late game compositions. Then there are maps. All maps are mirrored and have equal resources (for tournaments at least)
Anyway back to the sc2 wcs championship
What I really want to see with this kind of AI is it being developed and to a certain extent optimised on StarCraft 1 and then ported onto StarCraft 2. This is an easy transition for a human player but I imagine unbelievably difficult for a bot but it would be an amazing test of true machine learning and application.
sc1 and sc2 are very different, broodwar is way more micro intensive due to the limits on units in control groups on top of the changes in how the three races interact. SC2 meta can change a lot between patches let alone expansions. A big thing about broodwar to sc2 is that protoss in broodwar basically plays like terran in sc2, loads of units from broodwar don't exist in sc2.
A great many pro's just didn't transition.
The main thing is that the AI would have to play via the games ui and not directly into the api due to the apm and input lag of human beings. The AI program would have to be looking at the game so would need itself virtual keyboard and mouse that it has to use on an actual screen.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2025
