Transcript StarCraft Data Analysis

DATA
ANALYSIS
BACKGROUND
Real-Time Strategy(RTS) game released by
Blizzard Entertainment at 1998.
Three races: Terran, Zerg, Protoss
GAME TERMINOLOGY
Building
Unit
GAME TERMINOLOGY
Resources: Currency required in the game to buildings and
units. Two kind: mineral and gas.
Upgrade: Enhancements unit/building attack and defense
powers. Bought from different buildings.
Fog of War: Fog that prevents a player to see all the game
map
Line of Sight: Determines how much fog of war a unit/building
uncovers
Build: Strategy followed by a player(i.e chosen buildings, units,
how many of them)
Base: Concentration of buildings. Usually near mineral and gas
resources, with a main building at the center.
DATA SET
http://eis.ucsc.edu/StarCraft_Data_Mining
About four thousand collected game replay data. Collected
through StarCraft fan sites. Labeled by expert gameplayers.
Attributes: First building times of buildings and first training times
of units. Numerical values with interval scale.
Class attribute: Middle Game Build. Categorical. Seven for each
of three races.
AIMS
• Predicting middle game strategy real time
• Searching for gameplay patterns of professional
players
DATA MINING
‘UNKNOWN’ CLASS
MOST USED/UNUSED BUILDS
PROTOSS
Protoss vs Zerg
fast expansion -> %51,7
fast observer -> 1
Protoss vs Terran
fast observer -> %34,2 fast expansion -> %33,3
carrier -> %0,8
Protoss vs Protoss
fast observer -> 34,4
carrier -> %0,5
MOST USED/UNUSED BUILDS
TERRAN
Terran vs Zerg
Bio -> %88,4
Vulture Harass %0,5
Terran vs Protoss
Siege Expansion -> %23,8 Two Factory -> %20,3
Fast Dropship -> %8,9
Terran vs Terran
Vulture Harass -> %38 Two Factory %31,6
Standard %0,2
MOST USED/UNUSED BUILDS
ZERG
Zerg vs Terran
Three Hatch Mutalisk %31,8
Hydralisk Rush %1,3
Zerg vs Protoss
Hydralisk Mass %61
Lurker %3,8
Zerg vs Zerg
Two Hatch Mutalisk -> %92,5
Lurker -> %0,1 Standard -> %0,1
CLASSIFACTION
DECISION TREE
Weka J48 (C4.5) Decision Tree (%75 of file training, %25 testing)
Protoss vs Protoss (%71 correct classification)
Protoss vs Terran (%90 correct classification)
Protoss vs Zerg (%86 correct classification)
Terran vs Terran (%85 correct classification)
Terran vs Protoss (%88 correct classification)
Terran vs Zerg (%91 correct classification)
Zerg vs Zerg (%98 correct classification)
Zerg vs Protoss (%87 correct classification)
Zerg vs Terran (%86 correct classification)
CLASSIFICATION AMONG DIFFERENT
OPPONENTS
Protoss
Best Case: %73
Worst Case: %43
Terran
Best Case: %68
Worst Case: %22
Zerg
Best Case: %85
Worst Case: %38
RESULTS & COMMENTS
• Worst Cases happen when the opponent is Zerg. Over usage of
one strategy makes the data biased.
• Terran and Protoss change the execution style of strategies
against different opponents
• Zerg usually has similar style of strategy execution against
different opponents.
REAL TIME CLASSIFICATION
Results are generally good but useless in a real StarCraft game.
Mid Game Strategies are executed about 15th-20th minute of a
game
Players don’t have information on exact building times
Rebuild the data set as ordinal values by discretization(via
RapidMiner)
Don’t count values after 12th minute of the game
CLASSIFICATION AFTER REBUILDING
Weka J48 (C4.5) Decision Tree (%75 of file training, %25 testing)
Protoss vs Protoss (%80 correct classification)
Protoss vs Terran (%83 correct classification)
Protoss vs Zerg (%88 correct classification)
Terran vs Terran (%81 correct classification)
Terran vs Protoss (%74 correct classification)
Terran vs Zerg (%91 correct classification)
Zerg vs Zerg (%98 correct classification)
Zerg vs Protoss (%87 correct classification)
Zerg vs Terran (%86 correct classification)
FEATURE ELIMINATION
Trees are too complex to determine a rule
Features with very low number of
occurences are removed
Also features that doesn’t appear until 12th
are removed
Protoss
Build Pruned
Decision Tree
(C4.5)
Terran Build against Protoss
Zerg vs Terran
Build
Zerg vs Protoss
Pruned Decision Tree
PROFESSIONAL PLAYER PATTERNS
Tests with RIMARC:
Terran Two Factory Build
Terran Siege Expansion Build
PROFESSIONAL PLAYER PATTERNS
Terran Bio Build
Terran Fast Dropship
Build
USEFUL DATA OBTAINED




When the opponent is Zerg build is usually
fixed
For Protoss and Terran execution of
strategies differs according to opponent race
Rules to determine opponent build early in
game
Some patterns for different builds
THANKS FOR LISTENING
QUESTIONS..