StarCraft2 Replay analysis

data file

library(ggplot2)
library(corrplot)
library(corrgram)
library(gridExtra)

This dataset is an aggregate of the screen-fixations from screen movements of StarCraft 2 replay files. Quick look at the data file to detect (if any) outliers and/or NA values.

Questions worth exploring:

• How do the replay attributes differ by level of player expertise?
• What are significant predictors of a player’s league?
• it would be interesting to look at the LeagueIndex vs. the number of hours per week, or vs. players’s skills (ActionLatency for example)

df<-read.csv('../../SC2/starcraft.csv',sep=',')
str(df)

## 'data.frame':    3395 obs. of  21 variables:
##  $ GameID            : int  52 55 56 57 58 60 61 72 77 81 ...
##  $ LeagueIndex       : int  5 5 4 3 3 2 1 7 4 4 ...
##  $ Age               : int  27 23 30 19 32 27 21 17 20 18 ...
##  $ HoursPerWeek      : int  10 10 10 20 10 6 8 42 14 24 ...
##  $ TotalHours        : int  3000 5000 200 400 500 70 240 10000 2708 800 ...
##  $ APM               : num  144 129 70 108 123 ...
##  $ SelectByHotkeys   : num  0.00352 0.0033 0.0011 0.00103 0.00114 ...
##  $ AssignToHotkeys   : num  0.00022 0.000259 0.000336 0.000213 0.000327 ...
##  $ UniqueHotkeys     : num  5.49e-05 6.92e-05 4.19e-05 1.07e-05 3.85e-05 2.13e-05 6.74e-05 5.97e-05 1.88e-05 9.98e-05 ...
##  $ MinimapAttacks    : num  1.10e-04 2.94e-04 2.94e-04 5.33e-05 0.00 ...
##  $ MinimapRightClicks: num  0.000392 0.000432 0.000461 0.000543 0.001329 ...
##  $ NumberOfPACs      : num  0.00485 0.00431 0.00293 0.00378 0.00237 ...
##  $ GapBetweenPACs    : num  32.7 32.9 44.6 29.2 22.7 ...
##  $ ActionLatency     : num  40.9 42.3 75.4 53.7 62.1 ...
##  $ ActionsInPAC      : num  4.75 4.84 4.04 4.92 9.37 ...
##  $ TotalMapExplored  : num  0.00022 0.000381 0.000231 0.000202 0.000289 ...
##  $ WorkersMade       : num  0.001397 0.001193 0.000745 0.000426 0.001174 ...
##  $ UniqueUnitsMade   : num  4.71e-05 8.65e-05 6.29e-05 7.46e-05 7.70e-05 6.38e-05 5.62e-05 8.95e-05 6.58e-05 7.49e-05 ...
##  $ ComplexUnitsMade  : num  0 0 0 0 0 ...
##  $ ComplexAbilityUsed: num  0.00 2.08e-04 1.89e-04 3.84e-04 1.93e-05 ...
##  $ MaxTimeStamp      : int  127448 57812 95360 93852 51936 94032 89012 100556 106308 80136 ...

summary(df)

##      GameID       LeagueIndex         Age         HoursPerWeek   
##  Min.   :   52   Min.   :1.000   Min.   :16.00   Min.   :  0.00  
##  1st Qu.: 2464   1st Qu.:3.000   1st Qu.:19.00   1st Qu.:  8.00  
##  Median : 4874   Median :4.000   Median :21.00   Median : 12.00  
##  Mean   : 4805   Mean   :4.184   Mean   :21.65   Mean   : 15.91  
##  3rd Qu.: 7108   3rd Qu.:5.000   3rd Qu.:24.00   3rd Qu.: 20.00  
##  Max.   :10095   Max.   :8.000   Max.   :44.00   Max.   :168.00  
##                                  NA's   :55      NA's   :56      
##    TotalHours             APM         SelectByHotkeys   
##  Min.   :      3.0   Min.   : 22.06   Min.   :0.000000  
##  1st Qu.:    300.0   1st Qu.: 79.90   1st Qu.:0.001258  
##  Median :    500.0   Median :108.01   Median :0.002500  
##  Mean   :    960.4   Mean   :117.05   Mean   :0.004299  
##  3rd Qu.:    800.0   3rd Qu.:142.79   3rd Qu.:0.005133  
##  Max.   :1000000.0   Max.   :389.83   Max.   :0.043088  
##  NA's   :57                                             
##  AssignToHotkeys     UniqueHotkeys       MinimapAttacks     
##  Min.   :0.0000000   Min.   :0.000e+00   Min.   :0.000e+00  
##  1st Qu.:0.0002042   1st Qu.:3.275e-05   1st Qu.:0.000e+00  
##  Median :0.0003526   Median :5.340e-05   Median :3.990e-05  
##  Mean   :0.0003736   Mean   :5.873e-05   Mean   :9.831e-05  
##  3rd Qu.:0.0004988   3rd Qu.:7.865e-05   3rd Qu.:1.189e-04  
##  Max.   :0.0017522   Max.   :3.376e-04   Max.   :3.019e-03  
##                                                             
##  MinimapRightClicks   NumberOfPACs      GapBetweenPACs    ActionLatency   
##  Min.   :0.0000000   Min.   :0.000679   Min.   :  6.667   Min.   : 24.09  
##  1st Qu.:0.0001401   1st Qu.:0.002754   1st Qu.: 28.958   1st Qu.: 50.45  
##  Median :0.0002815   Median :0.003395   Median : 36.724   Median : 60.93  
##  Mean   :0.0003874   Mean   :0.003463   Mean   : 40.362   Mean   : 63.74  
##  3rd Qu.:0.0005141   3rd Qu.:0.004027   3rd Qu.: 48.291   3rd Qu.: 73.68  
##  Max.   :0.0040408   Max.   :0.007971   Max.   :237.143   Max.   :176.37  
##                                                                           
##   ActionsInPAC    TotalMapExplored     WorkersMade       
##  Min.   : 2.039   Min.   :0.0000913   Min.   :0.0000770  
##  1st Qu.: 4.273   1st Qu.:0.0002244   1st Qu.:0.0006830  
##  Median : 5.096   Median :0.0002695   Median :0.0009052  
##  Mean   : 5.273   Mean   :0.0002825   Mean   :0.0010317  
##  3rd Qu.: 6.034   3rd Qu.:0.0003253   3rd Qu.:0.0012587  
##  Max.   :18.558   Max.   :0.0008319   Max.   :0.0051493  
##                                                          
##  UniqueUnitsMade     ComplexUnitsMade    ComplexAbilityUsed 
##  Min.   :1.970e-05   Min.   :0.000e+00   Min.   :0.0000000  
##  1st Qu.:6.780e-05   1st Qu.:0.000e+00   1st Qu.:0.0000000  
##  Median :8.220e-05   Median :0.000e+00   Median :0.0000203  
##  Mean   :8.455e-05   Mean   :5.943e-05   Mean   :0.0001419  
##  3rd Qu.:9.860e-05   3rd Qu.:8.555e-05   3rd Qu.:0.0001814  
##  Max.   :2.019e-04   Max.   :9.023e-04   Max.   :0.0030837  
##                                                             
##   MaxTimeStamp   
##  Min.   : 25224  
##  1st Qu.: 60090  
##  Median : 81012  
##  Mean   : 83598  
##  3rd Qu.:102074  
##  Max.   :388032  
## 

anyNA(df)

## [1] TRUE

Comments

• Most of the features are numeric data.
• there is a maximum of 57 NA’s
• There is 1 outlier (TotalHours) : max is 100000 but the mean and median are much less (500, 960 respectively)

This outlier can be seen directly by plotting the distribution of TotalHours:

totHours<-ggplot(data=df,aes(x=TotalHours)) + geom_histogram(bins=100)
print(totHours)

## Warning: Removed 57 rows containing non-finite values (stat_bin).

Cleanup data

For a first look at the data I will simply remove the NA’s and the outlier. Later I will impute the data properly.

df <- na.omit(df)
df<-df[df$TotalHours!=1000000,]
totHoursCleaned<-ggplot(data=df,aes(x=TotalHours)) + geom_histogram(bins=100)
print(totHoursCleaned)

Correlation plot

Some attributes (features) may be correlated together so it’s worth looking at the correlation matrix.

num.cols <- sapply(df, is.numeric)
cor.data <- cor(df[,num.cols])
corrPLOT<-corrplot(cor.data,method='ellipse')

Definition League Index

league<-function(x){
    if (x==1) {return('Bronze')}
    else if (x==2) {return('Silver')}
    else if (x==3) {return('Gold')}
    else if (x==4) {return('Platinum')}
    else if (x==5) {return('Diamond')}
    else if (x==6) {return('Master')}
    else if (x==7) {return('Grand Master')}
    else{return('OTHER')}
}
df$LeagueName<-sapply(df$LeagueIndex,league)

Data exploration

Action Latency and Gap Between PACS show a very strong correlation so we can start to look at these features.

vs. Age

actionLatVsLeague<-ggplot(data=df,aes(x=factor(LeagueIndex),y= ActionLatency)) + 
geom_boxplot(aes(fill=factor(Age))) + theme(legend.position=c(.9, .65)) + 
xlab('League Index') + ylab('Action Latency')
print(actionLatVsLeague)

gapPacVsLeague<-ggplot(data=df,aes(x=factor(LeagueIndex),y= GapBetweenPACs)) + 
geom_boxplot(aes(fill=factor(Age))) + theme(legend.position=c(.9, .65)) + 
xlab('League Index') + ylab('Gap Between PACs')

print(gapPacVsLeague)

Comments :

• for a league Index increasing (= prestigious league), the latency (= time between clicks) is decreasing
• inside a given league, the latency is increasing with the Age
• same observations for GapBetweenPACs feature

Time spent playing

hourWeekVsLeague<-ggplot(data=df,aes(x=factor(LeagueIndex),y= HoursPerWeek)) + 
geom_boxplot(aes(fill=factor(Age))) + theme(legend.position=c(.9, .65)) + 
xlab('League Index') + ylab('HoursPerWeek')
print(hourWeekVsLeague)

totHoursVsLeague<-ggplot(data=df[df$TotalHours<5000,],aes(x=factor(LeagueIndex),y= TotalHours)) + 
geom_boxplot(aes(fill=factor(Age))) + theme(legend.position=c(.9, .65)) + 
xlab('League Index') + ylab('Total Hours')
print(totHoursVsLeague)

Comments :

• to expect to reach the higher league, you obviously need to play a lot, and regurlarly
• note that I have selected only the TotHours less than 5000 to have the data points in the plot frame
• there are few outliers () where the TotalHours is greater than 5000) :

##      GameID LeagueIndex Age HoursPerWeek TotalHours      APM
## 8        72           7  17           42      10000 212.6022
## 11       83           3  16           16       6000 153.8010
## 771    2246           5  22           16      20000 248.0490
## 1979   5610           4  22           10      18000 152.2374
## 2141   6020           5  22           10       9000 106.0056
## 2217   6242           3  24           20      10260  76.5852
## 2325   6518           6  20            8      25000 247.0164
## 3254   9055           3  19           20       6000 102.0114
##      SelectByHotkeys AssignToHotkeys UniqueHotkeys MinimapAttacks
## 8        0.009039739     0.000676240   0.000059700    0.001163531
## 11       0.001676615     0.000318557   0.000067100    0.000000000
## 771      0.023703208     0.000390712   0.000130237    0.000000000
## 1979     0.011983053     0.000205514   0.000015800    0.000015800
## 2141     0.003568862     0.000634731   0.000095800    0.000946108
## 2217     0.000779779     0.000196701   0.000000000    0.000063200
## 2325     0.015793839     0.000438389   0.000094800    0.000308057
## 3254     0.002045134     0.000317348   0.000044100    0.000044100
##      MinimapRightClicks NumberOfPACs GapBetweenPACs ActionLatency
## 8           0.001253033  0.004952464        24.6117       41.7671
## 11          0.000821541  0.003772383        23.4107       48.0711
## 771         0.000204659  0.004651336        37.8795       45.3760
## 1979        0.000363602  0.003351461        52.1896       63.9811
## 2141        0.000574850  0.003616766        28.6645       55.9603
## 2217        0.000316127  0.002437688        42.9480       84.6340
## 2325        0.001338863  0.004644550        17.6471       37.1837
## 3254        0.000555360  0.003032440        62.5423       67.3140
##      ActionsInPAC TotalMapExplored WorkersMade UniqueUnitsMade
## 8          6.6104      0.000447512  0.00227730     0.000089500
## 11         7.0044      0.000402387  0.00159280     0.000117363
## 771        4.7560      0.000390712  0.00152560     0.000111632
## 1979       4.9575      0.000300367  0.00067978     0.000079000
## 2141       4.6159      0.000299401  0.00101800     0.000095800
## 2217       5.9107      0.000189676  0.00044960     0.000070300
## 2325       6.5944      0.000343602  0.00186020     0.000071100
## 3254       6.3605      0.000211566  0.00141040     0.000052900
##      ComplexUnitsMade ComplexAbilityUsed MaxTimeStamp   LeagueName
## 8         0.000129281         0.00024862       100556 Grand Master
## 11        0.000000000         0.00001680        59644         Gold
## 771       0.000000000         0.00000000        53748      Diamond
## 1979      0.000000000         0.00000000        63256     Platinum
## 2141      0.000000000         0.00015569        83500      Diamond
## 2217      0.000245876         0.00035828       142348         Gold
## 2325      0.000000000         0.00001180        84400       Master
## 3254      0.000238011         0.00194820       113440         Gold

GamePlay

I will look at some features related to the gameplay of SC2 itself, meaning the number of units, special units made per timestamp :

• WorkersMade: Number of SCVs, drones, probes trained per timestamp
• UniqueUnitsMade: Unique units made per timestamp
• ComplexUnitsMade: Number of ghosts, investors, and high templars trained per timestamp

The plots are still vs the LeagueIndex to see the impact/experience of player’s gameplay within a given leauge

workersPerLeague<-ggplot(data=df,aes(x=factor(LeagueIndex),y= WorkersMade)) + 
geom_boxplot(aes(fill=factor(Age))) + theme(legend.position=c(.9, .65)) + 
xlab('League Index') + ylab('Workers Made')
print(workersPerLeague)

uniqueUnitsMadeVsLeague<-ggplot(data=df,aes(x=factor(LeagueIndex),y= UniqueUnitsMade)) + 
geom_boxplot(aes(fill=factor(Age))) + theme(legend.position=c(.9, .65)) + 
xlab('League Index') + ylab('Unique Units Made')
print(uniqueUnitsMadeVsLeague)

complexUnitsMadeVsLeague<-ggplot(data=df,aes(x=factor(LeagueIndex),y= ComplexUnitsMade)) + 
geom_boxplot(aes(fill=factor(Age))) + theme(legend.position=c(.9, .65)) + 
xlab('League Index') + ylab('Complex Units Made')
print(complexUnitsMadeVsLeague)

Comments :

• The number of workers per timestamp increase with the League Index, meaning higher League players can produce more units
• The number of Unique Unitsper timestamp looks constant vs the League Index ; it means that this may not be an important feature of the game difficulty (note that I’m not a SC2 player)
• The number of complex units per timestamp seems to increase with the League Index, meaning it may be important to be able to produce complex unit when you’re playing in higher difficulties