Approximation Models of 
Combat in StarCraft 2
Ian Helmke
Dan Kreymer
Karl Wiegand
CS 7180 - Sliva
October 21, 2012

Outline
Background and Motivation
Problem Statement
Approach and Method
Results
Applications and Limitations
Future Work

Background: StarCraft 2
Real-Time Strategy (RTS) computer game
Gameplay: Gather, build, research, attack, and then destroy all enemy buildings
Complex, zero-sum, partial information game
Popular game series for testing AI systems

Problem Statement: One Battle
Given the composition of two armies, which army will win and which units will remain?





Terran
Zerg

12 Marines

4 Marauders

4 Hellions

24 Zerglings

4 Roaches

4 Hydralisks
50%
50%

2 Marines

2 Marauders

1 Hellions


3 Hydralisks

Approach: Approximations
Units have health, DPS, flags, and basic attributes based on unique type

APX1: Randomized perfect focus fire
APX2: Free round of attacks by ranged units

APX3: Attributes and bonus damage

APX4: Preferred targeting of melee units

Method: In-Game Testing
12 matches: 4 rounds of 3 different combinations (PvT, TvZ, PvZ)
Each round increased complexity and size

12 custom maps with 10 battles per match

Tracked compositions and win percentages

APXn: 12 simulations based on 1,000 rounds

Example: Replay of Test Battle

Results: Predicting Victory

Results: Predicting Remaining Units

Results: Predicting Remaining Units

Applications and Limitations
Good:
Victory determination in decision trees
Heuristic for many planning algorithms
Selectively use/combine this model

Bad:
No micromanagement or formations

Okay:
No technology or spells

Future Work
Technology as a set of rules and effects
Spellcasters assumed to be fully effective
More testing or creation of a test corpus
POMDP for victory prediction
Combine with player-specific models to account for micromanagement

?
Full report, code, maps, replays, and all results are available at: [https://bitbucket.org/karlwiegand/sc2apx].