“StarCraft, one of the most popular games ever made, also serves as the perfect proving ground for artificial intelligence. Here’s the inside story of how a Berkeley team won the world’s first StarCraft AI competition with code that can beat even pro-level human players.”
Thanks for posting this, that’s a very interesting article. With game developers putting more and more emphasis on online play lately, this might motivate some of them to spend some time making smarter computer opponents for campaign games.
Such a bot could probably be tweaked to automatically adjust to the player’s skill level in order to keep the game challenging without being overwhelming. That’s definitely more fun than a computer that cheats to be more difficult.
Not surprising.
Starcraft is the game where you have to pay attention to at least ten things at once, and micro manage your units well in combat. This is exactly what computers can be way better at than humans. How can you play against an opponent with several thousands of APM? Have a look at the mutalisks vs. high templars battle on the linked page; it’s scary.
This is not to say that their achievement is not impressive. It is. However, they have achieved something (beating a strong amateur) in 1.5 years. A similar feat took way longer for chess, not to mention Go, where it has only been done very recently.
What exactly is new here?
We’ve had bots for years now… any1 remember the Quake Bots? Thats where the whole ‘bot-business’ (yeye, u can have a laugh at that!
) took off, and those bots ALSO had to take care of several things at the same time (so-called multitasking)
Whats interesting is the amount of effort it takes to emulate (well, really, its simulate, but you can ponder the difference
) a human player!
You are completely missing the point. Its one thing to create a fps bot where reflexes count as much as anything. There, the computer has a huge advantage. Its quite another thing to build a bot that plays a strategy game where quickness counts for very little. Not to mention, this AI had three main tasks (outlined in the article) not just combat like a fps. This is quite an impressive showing. Also note, that while the short term goal is to beat a human, the real competition is beating the other AI’s.
Yup. FPS AIs are relatively straightforward – one unit to manage, shoot at whatever moves, and if nothing to shoot, go and find something. It might also handle simple objectives like “guard the base” or “fetch the enemy flag”.
An RTS AI needs to do that for hundreds of units of differing capabilities, and make them work as a team – e.g infantry need to engage the enemy without getting pulped by their own artillery and air support. The RTS AI also needs to produce all those units, working out which ones it needs (or can afford), and needs to obtain the resources to build them with. It’s a much higher level of complexity than the FPS AI…
Hmm.. reading the article in more detail, it looks like having recognised those difficulties, they’ve chosen to avoid them as much as possible, rather than solve them – focusing on one versatile unit so as to simplify teamwork and build management…
Overmind – so that’s what Nietzsche was talkiing about.