Electronic sport, or e-sport, denotes the extreme practice of video games where so-called cyber-athletes compete in world-wide tournaments. As for any sport, such professionals are surrounded by sponsors and practice within professional teams. These professional games are even broadcast by commentators over specialized TV channels. {\sc StarCraft II} (Blizzard Entertainment) is one of the most competitive video game and has now its own world-wide players ranking system (based on the well-known ELO system) and annual world cup competition series (WCS) with a US\$1.6 millions prize pool for the year 2013. Each match between two opponents can be recorded as an exhaustive and noisy sequence of actions. Analyzing these sequences yields an important outcome for game strategy prediction and in-depth game understanding. In this work we report a preliminary study on StarCraft II professional players strategies' discovery based on sequential pattern mining.

1. An#approach#with#sequen1al#pa3ern#mining#
!
G.!Bosc,!M.!Kaytoue,!C.!Raïssi,!J.5F.!Boulicaut!
Machine!Learning!and!Data!Mining!for!Sports!AnalyCcs,!MLSA@ECML/PKDD!2013!workshop,!27!September!2013,!Praha,!Czech!Republic!

2. o !Anyone,'any)me,'anywhere,'…'any$what!
o !Flourishing'video'game'industry!
O WORLD!OF!WARCRAFT!(2004)!aXracted!millions!of!gamers,!US!20$/month/user!
O GRAND!THEFT!AUTO!!V!(2013)!:!!US!200!M$!budget,!US!800!M$!sells!in!24h!
O CANDY!CRUSH!SAGA:!45!M!users,!15!M!users!acCve!every!days!
O WATCH!DOGS!(2013),!and!many!others,!being!adapted!to!cinema!!
o !A!context!revealing!the!former!niched!compe))ve'gaming'
2'

3. A!new!markeCng!landscape!
!
Entering!in!popular!culture!
!
A!sport?!!

4. o Professional'&'amateurs!(coaches,!teams,!fanclubs,!streaming)!
o Tournaments!(commentators,!sponsors)!
$BLIZZARD$WCS$2013:#$1.6#million#USD#
#DOTA$2#The$interna<onal$Aug.$2013:$$2.2#millions#USD#
o AXracCng!industrials,!sponsors!:!
5!Sept.!2013:!Samsung!in!the!process!
to!form!a!Pro!Team!!(300~500K!$)!
5!Intel,!Nvidia,!…!community!managers!
o Followed!in!great!numbers,!more!and!more!are!making!a!living!of!it!
G.!Bosc!et!al.!5!Machine!Learning!and!Data!Mining!for!Sports!AnalyCcs,!ECML/PKDD!2013!workshop,!27!
September!2013,!Prague,!Czech!Republic! 4'

5. Main!Principes!of!RTS!games!
Build!orders!openings!
Strategy!balancing!!

6. o A!map,!aerial!view!(like!gmap)!
o 2'players!in!command!(1v1)!
o Controls!units!and!building!
o Economy'Gather!resources!
o Produc)on'and'technology'
Construct!new!buildings!
o Training!combat!units!from!buildings!
o Order!to!explore!the!map!
!!!(fog!of!war)!
o Order!to!aBack!an!enemy!
o Winning'condi)on':!opponent!
resigns!or!has!no!more!forces!!
6'

7. ! Can!operate!fast'mul)Fdecisions'making'process'in'uncertain'
environment'
! Mul)Ftasking'(macro,!micro,!several!fronts,!Cmings,!scouCng…)!
! Velocity!(200!to!400!clicks/Key!pressed!per!minute)!
! Decisions'!Rock/Paper/Scissors!principle,!army!composiCon,!
aXacks,!…!
! Uncertainty'under!a!fog!of!war!
! Ac)ons'may'fail'misclicks!
'
It'requires'strategies,'training,'metaFgaming,'etc.'
7'
AutomaCc!discovery!of!build!orders!openings!!
A!challenge!for!game!editors!and!(pro)5gamers!

8. Data'as'traces'of'interac)on'
'
Sequen)al'paBern'mining''
'
Build'order'opening'discovery''
'
Balance'of''an'opening'

9. 9'
! A!series!of!)mestamped!acCons!of!different!types!for!both!players!

10. • 'Sequence!a!series!of!events!from!an!alphabet!
• !Sequence'database!a!set!of!sequences!
• 'Frequent'sequen)al'paBern'an!arbitrary!
sequence!more!general!than!(supported!by)!a!
given!proporCon!of!the!database!
• E.g.!sejng!minimal!support!to!75%!
• <acc>!is!frequent!
• <a{bc}a>!is!not!frequent!
10'
id' sequence'
1! <a{abc}{ac}d{cf}>!
2! <{ad}c{bc}{ae}>!
3! <{ef}{ab}{df}cb>!
4! <eg{af}cbc>!

11. • !How'to'turn'replays'files'to'sequences!!
• Events?!
• Windows!of!Cme?!
• One!or!two!players?!
• !How'to'validate'extracted'paBerns''
• Measuring!the!balance!of!the!strategy?!
• !How'to'analyze'the'results'
• Search!&!Sort!&!Filter!
• PaXerns!to!be!re5used!for!various!task!
SC2!replays!
Sequence!
Database!
Frequent!
PaXerns!
Openings!
balance!
11'

12. ! A!sequence!is!a!series'of'windows'of!)me'composed'of'ac)ons'done!by!both!players!
To!differenCate!players,!events!are!tagged!as!winner'(success)'or'looser'(fail)'(1v1!situaCon)!
! Events/Ac)ons!!are!elements!of!
12'

13. ! Given!a!frequent!sequenCal!paXern!extract!with!a!FPM!algorithm!
! !its!dual!represent!the!reversed!scenario,!where!the!!winner!would!have!lost!
! And!then!
13'

14. • !Takes'values'in'
• !Equilibrium'
• !UnFbalanced'strategy'
!
• 'Mirror'strategy'
14'

15. ! A'naïve'approach'
o Mining'step:!Extract!all!frequent!sequenCal!
paXerns!with!a!known!algorithm!(e.g.!PrefixSpan)!
o PostFprocessing'step:!count!how!many!Cmes!the!
dual!of!each!paXern!appears!in!the!base!
!!!!!!!!!!!!!!!!Not!!scalable !!
! A'more'elaborated'approach'
o Mining'step:!Extract!all!frequent!sequenCal!
paXerns!with!a!know!algorithm!(e.g.!PrefixSpan)!
o PostFprocessing'step:!Take!advantage!of!the!tree!
data5structure!of!the!extracted!paXerns!
!Faster!computaCon!of!the!balance!
!Avoid!redundant!paXerns!
15'
More$technical$and$algorithmic$details$in$the$proceedings$

16. Dataset!
QuanCtaCve!results!
QualitaCve!results!!

17. o !RestricCng!371$267$SC2'replays'harvested!on!several!dedicated!websites!
o !90'768'games'(1v1!only,!APM>200,!Master/Grandmaster)!
G.!Bosc!et!al.!5!Machine!Learning!and!Data!Mining!for!Sports!AnalyCcs,!ECML/PKDD!2013!workshop,!27!
September!2013,!Prague,!Czech!Republic! 17'
APM
Time
(min)
Count replays
Time (min)
% Actions
Time (min)

18. o One!!sequence!database!for!each!matchFup!(PvsP,!PvsT,!PvsZ,!TvsZ,!TvsT,!ZvsZ)!
o ‘buildForders’!acCons!
o Windows!of!Cme!set!to!30!seconds!
18'

19. 19'

20. '
Top'frequent'paBerns'are'more'balanced''
They'should'represent'balanced'and'wellFknown'openings' 20'

21. Well'known'openings,'broken'strategies,'unFdiversity'in'mirror'matchups,'etc.'
'
21'
Forge!Expand!
• 9:!Pylon!
• 14:!Forge!
• 17:!Nexus!
• 17:!Pylon!
• 18:!Gateway!
• 18:!Photon!Cannon!
• 18:!Assimilator!x2!
15151!Build!
• 12:!Barracks!
• 13:!Refinery!
• 17:!Refinery!
• 18:!Factory!
• 20:!TechbLab!
• 22:!Starport!
4!gates!
• 9:!Pylon!
• 12:!Gateway!
• 14:!Assimilator!
• 15:!Pylon!
• 17:!CyberneCcs!pCore!
• 22:!Warpgate!
• 24:!Gateway!px3!
• 26:!Pylon!!
• 26:!Assimilator!
minSupp:#5%#
3148#pa3erns#
balance(s)=0.59$
minSupp:#5%#
400#pa3erns#
balance(s)=0.46$
minSupp:#5%#
591#pa3erns#
balance(s)=0.52$
Leveraged'by'
Blizzard!
(Queen!range+,!
Supply!before!
barracks)!
!
balance(s)#<#0.6$
balance(s)#>#0.6$
$
!
!

23. DataFmining'
o Emerging$paNern$mining:#items!are!labelled,!not!the!sequence!!itself!
o Scalability:#CompuCng!the!balance!measure!double!the!execuCon!Cme!in!worst!cases!
o Other$applica<ons:#two!agents!compete!for!a!resource,!one!succeed!while!the!other!fails!!
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!(sport!analyCcs,!AI,!planning,!…)!
Video'game'analy)cs'and'EFSport'
o Balance!is!one!of!the!hoXest!issue!in!the!E5Sport!scene!
o Service#that!can!help!pro5gamers!and!their!team!to!prepare!match!(e.g.!azer!a!SQL!query!on!replays)!
o Non!professional!entertainment'
Perspec)ves'
o PaXerns!to!be!re5used!as!‘knowledge!unit’!for!other!tasks,!e.g.!real5Cme!winning!predicCon!
o CompuCng!balance!during!the!mining!step!and!improving!specific!pruning!strategies!
o New!dimensions!to!characterize!emerging!paXerns:!!
This#strategy#is#unbalanced#in#map#x#during#season#y#for#player#z#
23'

24. An#approach#with#sequen1al#pa3ern#mining#
!
G.!Bosc,!M.!Kaytoue,!C.!Raïssi,!J.5F.!Boulicaut!
Machine!Learning!and!Data!Mining!for!Sports!AnalyCcs,!MLSA@ECML/PKDD!2013!workshop,!27!September!2013,!Praha,!Czech!Republic!
Source!code,!dataset!and!quesCons!
hXp://liris.cnrs.fr/mehdi.kaytoue!