AIIDE 2012 Feedback

[heavy_metal]

One thing that would improve interest and participation in future competitions is more publicity. For instance, when I search 'Starcraft AI' in google most of the results are for the 2010 comp and Overmind. The most recent article I see on the front page is about Matt's Starcraft 2 AI. Somehow 2010's Overmind and this guy's Starcraft 2 AI got onto a bunch of news sites. If AIIDE 2012 got some coverage it would be great.

[Taranok]

I think the focus needs to be more towards strategy. Right now, basically microing units and constantly building the same units, is what most AI's seem to do to win. To implement more strategy, I think putting more experimental maps in would be good idea. Maps like Chain Reaction, Arkanoid, 815, could encourage some more thinking. Maybe more maps with island expos too.

[Equalizer]

I think the key issue is that there really doesn't appear to be a starting framework to create an AI that can do what is need to be able to play strategy games.

Just getting an AI to play like the default scripted computer player isn't that easy (moving units together and having half decent building placement).


Also taking standard AI algorithms and using them to try and create an AI for Starcraft doesn't seem to work out very well (uncertainty, continuous/huge state space and fast time).

Flexible strategy seems more like an algorithmic problem (that the tools needed don't exist and they need to be made) rather than an implementation problem (using existing tools to solve the task) and is probably why it appears lacking cross the board.


The "where do I even start?" issue is at least why I haven't gotten anywhere.

[krasi0]

There are many good bots which are open source. Some, like UAlbertabot have good docs to start with... This is even better than a framework - you have a working solution to build on.
[quote name='Equalizer' date='05 October 2012 - 06:55 AM' timestamp='1349409342' post='11189']
The "where do I even start?" issue is at least why I haven't gotten anywhere.
[/quote]

[heavy_metal]

[quote name='Equalizer' date='04 October 2012 - 11:55 PM' timestamp='1349409342' post='11189']
Also taking standard AI algorithms and using them to try and create an AI for Starcraft doesn't seem to work out very well (uncertainty, continuous/huge state space and fast time).
[/quote]

It would be necessary to greatly simplify the state space by ignoring variables, making assumptions. Did you look at UAlbertaBot's micro search? It appears to be an application of classic game AI to a simplified Starcraft combat system.

[Equalizer]

Fair enough. Perhaps it would be useful to put together a form of resources list.

The closest thing to this that I known of is the "Projects Using BWAPI" on BWAPI google code site.

[Equalizer]

[quote name='heavy_metal' date='05 October 2012 - 08:01 PM' timestamp='1349492486' post='11191']
It would be necessary to greatly simplify the state space by ignoring variables, making assumptions. Did you look at UAlbertaBot's micro search? It appears to be an application of classic game AI to a simplified Starcraft combat system.
[/quote]

I did look at it a while back (ref. post on teamliquid) but doesn't that have scaling issues past a small number (~10) units?

Maybe there are other improvements since then.

[DaveChurchill]

[quote name='Equalizer' date='05 October 2012 - 07:31 PM' timestamp='1349494269' post='11193']
[quote name='heavy_metal' date='05 October 2012 - 08:01 PM' timestamp='1349492486' post='11191']
It would be necessary to greatly simplify the state space by ignoring variables, making assumptions. Did you look at UAlbertaBot's micro search? It appears to be an application of classic game AI to a simplified Starcraft combat system.
[/quote]

I did look at it a while back (ref. post on teamliquid) but doesn't that have scaling issues past a small number (~10) units?

Maybe there are other improvements since then.
[/quote]

The micro search for unit control is up to 8v8 but the battle simulator can handle any size battles