Ideas for a masters research paper on SpringRTS AIs?

[PicassoCT]

We.. and well the whole gameindustry could use a bot that balances units ingame. Its a tough topic though. But if you cut out tons of human balancetesters, the industry certainly will gild your dick with goldleafs.

[Tim Blokdijk]

That's actually a cool idea. An algorithm to balance units.
Something along the lines that you set some fixed characteristics (the role) that makes that unit unique and the algorithm balances the values relative to the other units.
Then you take a supercomputer, let it run 5000 AI v/s AI test games across 10 distinct maps and see how things add up.

So you would set up for example a 'scout' role and define some metrics that measure effectiveness. I presume that how many units it discovers relative to other units is a metric. And the scout role would also need some AI usage rules that match a 'scout' in action. As in "move to void while keeping some distance from enemy units that can kill the scout".

Maybe use a rule engine like http://clipsrules.sourceforge.net/?

The fun thing about it is that it's all quite realistic. This can actually work, as in.. if you don't do this I will do it (within the next 10 years).

[Licho]

Balancing units in this way is impossible.

You need to take human into equaiton. Units dont just exist in virtual space.

Stuff like perception, reaction times needed and similar affect how effective the unit is in real games.

For example if you make flak cloud less visible, it suddenly makes flak a lot more effective - simple perceptual change.

If you make unit that needs perfect reflexes it will be very strong in hands of AI or fast humans but crappy otherwise.


In ZK we measure real performance of units in hands of people in real games to evaluate them, but even this approach has limits.

[Petah]

Welcome back buddy.

I've been also meaning to polish of BAI (which I was making around the same time you where active) and make work with the latest interface.

[Zydox]

I've been also meaning to polish of BAI (which I was making around the same time you where active) and make work with the latest interface.

Start polishing, I wouldn't mind playing against a "new" AI

[PicassoCT]

Balancing units in this way is impossible.

You need to take human into equaiton. Units dont just exist in virtual space.

Stuff like perception, reaction times needed and similar affect how effective the unit is in real games.

For example if you make flak cloud less visible, it suddenly makes flak a lot more effective - simple perceptual change.

If you make unit that needs perfect reflexes it will be very strong in hands of AI or fast humans but crappy otherwise.


In ZK we measure real performance of units in hands of people in real games to evaluate them, but even this approach has limits.

There is actually a way to estimate even this (percentage on the screen) and the movie industry has done a whole century of try and error research, as to how long people expect fights to be. Also, you could exclude specialists, and auto-balance only the warriors.

A Algo that takes human interaction costs into account, could even bring units back into the game balancewise, that are usually "balanced out of the game" (e.g. airtransports) because the complexity to comand them, plus the risk is usually not taking into acount.

(You have to make the damn things flying fortresses to make them itnersting for players, only there passengers really on the damagelist)

This is acutally a project i would embark upon. if i hadnt signed myself into a lot of stuff allready (journeywar, cheeselobby, a RepRap). Im not a miracle man.
This would be a sweet miracle though. But i guess it would raise to many hungry eyes in the industry. Better to live poor in peace, then to dwell rich under siege.

[Licho]

I fail to see whats advantage of such "algo" compared to actually measuring real performance of units in hands of real people..

Obviously the first is just an imperfect attempt at aproximation of the latter.

[PicassoCT]

The month and years of testgames nessecary to balance out something like starcraft could be avoided by a learning algo, able to find local optimums of exictment, unitpower, avoiding long drawn fights like in SINS and forcing you to get rid of useless units.

[gajop]

Hey, nice to see you back again.
I too am a masters student in AI and Machine Learning now, although I probably won't be picking anything related to SpringRTS.

I did a bit of work on your springgrid/ailadder projects, mostly by merging it both to one project due to performance reasons, adding some missing features and migrating it to pylons to maintain sanity (it can be seen here: https://github.com/gajop/springgrid/tree/pylons and https://github.com/gajop/botrunner). Sadly I never got to finish it, but I may eventually return to it, and would certainly offer help if you wish to use this version.

If I may suggest a topic of research related to springrts, you could try to mine over existing player vs player replays, and try to collect existing build orders, and see which build order defeats which other build order. It would probably be a combination of Sequence Mining -http://en.wikipedia.org/wiki/Sequence_mining (to detect which build orders are occurring often enough) and a supervised learning algorithm with 3 inputs: first and second player build order sequences and a label indicating the winner.

Try to simplify it and only consider buildings, perhaps even on only one map and race, and game type (1v1 f.e). You could eventually do multiple tests for other maps, mods and races.

[PicassoCT]

dsd is so finnished when this bot leaves school