8-Button StarCraft lets the AI take care of the complexity | Hacker News Reader

8-Button StarCraft lets the AI take care of the complexity

bgweber.com

49 pointstantalor24 comments

Not sure how good this AI would be. It pretty much delegates micromanagement and lets the user make high-level decisions. However, it doesn't seem fine grained enough to handle different situations. The fact that the 8 buttons include unit names is slightly worrying; is this is specifically for Protoss? What if a reaver drop is the current optimal plan of attack? What about scouting? There were no scout units at all. The video just shows the person clicking "attack" and the zealots/dragoons automatically moving towards the enemy base, so maybe the "black sheep wall" cheat is presumed to be enabled. Scouting is one of the most important tactics. How would the system deal with cloaked units, etc? It just seems like a helper system that does trivial things like executing a standard build order, assigning probes to resources, etc.

I don't think this is intended to be a serious way of playing the game.

Yah this is kinda cool, but hardly practical in terms of any real game play.

Also, I think the beauty of starcraft is that it is extremely hard to master due to the complexity involved. Like chess, there are standard opening moves but beyond the initial sequences, the possibilities are almost endless.

I would like to see someone program a computer who could beat the best starcraft player. I bet that would make the IBM Deep Blue seem like a joke.

>>I would like to see someone program a computer who could beat the best starcraft player. I bet that would make the IBM Deep Blue seem like a joke.

Deep Blue is ancient, modern chess programs make it look like a joke ;-)

In SC a program can abuse early some insane micromanagement to win before any hard strategy/positional decisions have to be made.

Not true.

There is a trade off between speed and economy when doing an early rush. The attacking player will always have less attacking units than the defensive player due to the travel distance between bases. (Assuming the defensive player is prepared for the rush)

In addition, by the time any first tier attacking units reach you, it's always possible to have a defensive structure, which can deal fairly well with any sort of micromangement.

>> There is a trade off between speed and economy when doing an early rush. The attacking player will always have less attacking units than the defensive player due to the travel distance between bases. (Assuming the defensive player is prepared for the rush)

The attacking AI can abuse it's much much higher APM to do a lot of damage with less units. Like killing workers with zealots, zerglings or vultures.

>>In addition, by the time any first tier attacking units reach you, it's always possible to have a defensive structure, which can deal fairly well with any sort of micromangement.

Defensive structures are 1. expensive and 2. can be bypassed. An AI can use the same resources, human spent on defense, to expand or tech, easily winning the game.

I can definitely see better micro giving the AI an advantage, but it seems like it'd have to have at least competent strategic play to still do well; it's hard to recover from completely screwed-up build order. It'll be interesting to see how it plays out, though; if bots in the Starcraft AI competition (http://eis.ucsc.edu/StarCraftAICompetition) do start winning exclusively through insane micro and a hardcoded rush in the first 2 minutes, it might need to be changed to stay interesting.

To me, the most interesting part is seeing what kinds of AI architectures can simultaneously deal with real-time action across all the levels of complexity--- micromanaging individual units while also making unit-group-movement decisions, build-order decisions, expansion decisions, teching-up decisions, etc., which coordinate with each other. That might need a different game or a variant, though; it's possible that Starcraft is now too much like chess, where people have already mapped out giant databases of opening books (you can find detailed information on various SC wikis about exactly when to build different things in which order, what to build to counter other strategies, etc.). So it might be that a computer perfectly executing one of those hardcoded strategies with perfect micro is too close to optimal to be interesting. But I'm not sure that's true yet; current bots are actually not too great.

Alternately, someone could make an APM-limited variant to handicap the computer a little closer to human clicking speed. That would introduce a strange optimization problem where the AI would need to optimize its click usage as a limited resource.

Just wanted to point out that it's pretty silly to talk about an AI's APM, since its APM is infinite (bounded by hardware limits like clock speed, etc).

Just wanted to point out that it's pretty silly to say that infinite = bounded ;-)

I guess what I meant was that it would make the hype that Deep Blue got at the time seem like a joke, since I believe Starcraft is an immensely more complicated game.

And I disagree that insane micromanagement will win the game. In the opening moves, there's only so much micromanagement you can do since you're limited on resources , technology, and units. And by the time where lots of micromanagement comes into play, there are already a lot of strategic decisions made that could easily affect the outcome of the game.

Reminds me of the approach Master of Orion 3 took, sort of out of necessity: they found they had built a game with such absurd micromanagement that they had to add some AI managers to do some of the lowest-level "playing" for you in order to make the game actually interesting/enjoyable.

They failed miserably as I recall. It was a glorified screen saver.

Why is everyone bashing the AI? Its clearly a joke from the first two sentences of the post:

"John Davison recently claimed that games are "Too Big and Too Hard". I have addressed this issue by reducing StarCraft to 8 buttons."

Or is it too early in the morning for satire?

"Ben Weber is a Ph.D. candidate working with Michael Mateas in the Expressive Intelligence Studio at the University of California, Santa Cruz."

Heh, small world. Michael Mateas was my CS2260 professor at Georgia Tech and I was in the last class he taught before Michael left for Santa Cruz. For those wondering, 2260 was Media Device Architectures, a class based around coding for the GameBoy Advance. Here was part of the game I made for our final project: http://www.flickr.com/photos/pauls/133005425/

course: http://www.cc.gatech.edu/classes/AY2006/cs2260_spring/

The point of playing starcraft is to actually do all that "boring" macromanagement staff. That is a big bulk of the game, and certainly what separates players of different skill levels.

As long as you're abstracting out the human interaction, why not go whole hog and let the AI decide when to click those 8 buttons for you?

You can replace them all with one big button labeled "Play Starcraft", then sit back and watch it win for you. Like this:

http://progressquest.com/

Progress Quest is great for playing in work. I "played" it for months!

From the article: "In the fully realized system, EISBot, the AI decides which buttons need to be pressed."

The really cool thing here is if you click a few links in the article, you find an API to write your own Starcraft AI, and a competition for them.

This is the ultimate programming game!

http://code.google.com/p/bwapi/

Wow, someone built the system I was advocating for 3 months ago: http://news.ycombinator.com/item?id=1495760

Awesome.

It's rather interesting that he chose the faction requiring the least micromanagement, at least as far as base-building. Seems like a pretty neat tech demo, though.

So this AI tries to micromanage for you. But I could not get the game play here. You build a building and put it into a "build unit for me" mode. But what about the "attack mode?" why should AI decide where and when to attack? Doesn't this make game play just some sort of child play or a "logistics" game? I supply buildings and units, and AI decides the real strategy?

Hard to read in the video -- the buttons are:

The expand button is my favorite...

Except maybe defend, I think every single one of these would lose to Cannon -> DT rush.

Reply on news.ycombinator.com