Another update from the world of academic StarCraft AI! Here is a video demonstration of the results from our paper:
"Fast Heuristic Search for RTS Game Combat Scenarios"
We use a new algorithm (ABCD - Alpha Beta Considering Durations) to plan all unit actions you see being performed. The search is allowed to run for 50ms per frame in order to simulate real-time performance. There is *NO* human input or hard-coded scripting being used. All actions are being decided intelligently in a similar manner to how chess AI systems work. The reason this problem has been so hard is two fold:
- Extremely high branching factor for search (ie: an exponential number of possible simultaneous actions)
- Limited resource budget (50ms per decision, vs. many seconds or even minutes given to a game like chess)
Unit HP bars are shown above the units, as well as larger versions in the top-right corner. The red units are controlled by search, while the blue units are controlled by the default "attack closest unit" behaviour which is default to the StarCraft built-in AI. Initial starting positions for units are generated randomly, but symmetrically to guarantee fairness.
Shown is our combat simulation visualizer. Due to some issues with unknown unit behaviour in StarCraft, it is not yet possible to carry out perfect action execution in the game itself, so we showcase its performance first in simulation. Soon we'll be incorporating this into our AI bot "UAlbertaBot" which came 2nd place last year at the 2011 [AIIDE StarCraft AI Competition](http://www.starcraftaicompetition.com).
- Why is this new?
This is new because every other micro system so far (including those in BW, SC2, and the AI competition) has been performed using hard-coded human knowledge, or with incredibly long rule-based scripts. With this system, I can simply throw any units at it and it figures out the best moves on its own.
- Why does the script win sometimes?
In the examples I recorded, the blue player (script) does win sometimes. This is due to the nature of the randomly generated positions, as well as the visualization eating up some of our processing time. In our paper, our method beats the best scripts used from the 2011 StarCraft AI Competition over 80% of the time.