22
2016 IEEE CIG StarCraft AI Competition ( https ://sites.google.com/site/starcraftaic / ) Cheong-mok Bae, Seonghoon Yoon, Seok Min Hong Kyung-Joong Kim Cognition and Intelligence Lab (http://cilab.sejong.ac.kr) Sejong University, Seoul Republic of Korea

Starcraft 2016

Embed Size (px)

Citation preview

2016 IEEE CIG StarCraft AI Competition( https://sites.google.com/site/starcraftaic/ )Cheong-mok Bae, Seonghoon Yoon, Seok Min HongKyung-Joong Kim

Cognition and Intelligence Lab (http://cilab.sejong.ac.kr)Sejong University, SeoulRepublic of Korea

1

StarCraft Brood War Real-time strategy game Released in 1998 by BlizzardThree races Protoss, Terran, ZergPlay timeSeveral minutes ~ Hours Goal of the gameEliminate all the buildings of opponents

ResourceManagementUncertaintyHandlingMicroManagementDecision MakingReal-TimeResponse

StarCraft AI Competition Annual event since 2010 Three independent competitions (IEEE CIG, AIIDE, and SSCAIT)C++ or JAVA programming using BWAPI (Brood War API)AI vs. AI matches Fully automated tournament software 8 days running with 17 machines

Basic Rules

AI1AI2AI4AI5AI3

Full round-robin style tournament100 Rounds with 5 maps File I/O is allowed Used for long-term adaptation 42 milliseconds response time requirement 1 hour time limit for each match

4

2016 Update on RulesOpen source policy All participants source code will be available after this competition

Introducing qualifying and final tournaments Qualifying Select top players (50%) to advance 2nd stage FinalFull round-robin tournaments of the top players Winner is determined from the 2nd Stage

Qualification(All entries)Final(Half)

Number of Games Played

[Source] 2010~2015 data from https://webdocs.cs.ualberta.ca/~cdavid/starcraftaicomp/report2015.shtml

6

Entrants16 participants from 13 countries9 Updated or newcomer submissions7 Last years bot without change (Re-entrance)

AIUR Bonjwa OpprimoBot Overkill Salsa TerranUAB UAlbertaBotLast Years bot(Re-entrance)LetaBot MegaBot SRbotOne XelnagaII ZiaBot Iron tscmoo Tyr ZZZKBot Newcomers & Updated

7

Entrants (Details)

AiurBonjwaOpprimoBotOverkillSalsaTerranUABUAIbertaBotLetaBotMegaBotSRbotOneXelnagaIIZiaBot (Navinad)IrontscmooTyrZZZKBotProtossFlorian RichouxDustin DannenhauerJohan HagelbackSijia XuPablo Garcia SanchezFilip BoberDavid ChurchillMartin RooijackersAnderson Rocha TavaresJohan KayserCheong-mok BaeSungguk ChaIgor DimitrijevicVegard MellaSimon PrinsChris CoxeTerranTerranZergZergTerranRandomTerranProtossTerranProtossZergTerranTerranTerranZergUniversit de NantesLehigh UniversityLinnaeus University-University of GranadaPoznan Univ. of TechnologyUniversity of AlbertaMaastricht UniversityUniversidade Federal de Minas UQAMSejong UniversityUNIST--Utrecht University--------StudentStudentStudentStudent---ProgrammerSW Engineer/Developer

8

Race Distribution

12345678902013201420152016

Zerg

Terran

Protoss

Random

9

Results Announcement

Result Qualifying Stage 11 games Lost : Network Problem, and etc.Last 1 game didnt run: Tournament Manager BugBotIronTscmoo LetaBotOverkillMegaBotUAlbertaBotZZZKBotAiurTyrZiabotTerranUABSRbotOneOpprimoBotXelnagaIIBonjwaSalsaRaceGames1500149815001499149914991499149814991497149814991498149714991497TerranTerranTerranZergProtossRandomZergProtossTerranZergTerranTerranTerranProtossTerranZergWin118811531111106410511038103794692469550233233131028422Loss31234538943544846146255257580299611671167118712151475Win%79.276.9774.0770.9870.1169.2569.1863.1561.6446.4333.5122.1522.120.7118.951.47AvgTime14:1114:1513:3411:3711:449:597:2313:2017:199:5614:1114:4516:4613:3412:0510:53Hour3083816183003277211819051320Crash22271522102007251837232046771480Timeout7000400001141602260119Rank12345678910111213141516--119881198811988-12:511951840477Total

11

12

Result Final StageBotTscmooIronLetaBotZZZKBotOverkillUAlbertaBotMegaBotAiurRaceGames700700700699700699700700TerranTerranZergRandomProtossWin456381376371360343266246Loss244319324328340356434454Win%65.1454.4353.7153.0851.4349.0738.0035.14AvgTime15:2915:2313:458:4213:1611:2210:5612:54Hour317811522515Crash01400401002Timeout03000160Rank12345678--279927992799-12:434312010TotalTerranZergProtoss

Congratulation!Winner of 2016 IEEE CIG StarCraft AI Competition TSCMOO by Vegard Mella from Norway

13

Rank Change

BotIronTscmoo LetaBotOverkillMegaBotUAlbertaBotZZZKBotAiurRank12345678Win%79.276.9774.0770.9870.1169.2569.1863.15QualificationBotTscmooIronLetaBotZZZKBotOverkillUAlbertaBotMegaBotAiurRank12345678Win%65.1454.4353.7153.0851.4349.0738.0035.14Final

15

Winner Tscmoo

Comeback with Terran Bot (Last year: Zerg)Its good with vultures Use neural network Long Short-Term Memory (LSTM) for some high level decisions (which units to build) to provide hints to bunch of heuristics Remember what happen in previous games, and do something reasonable in response next game

Terran

2nd Rank Iron

A successor of Stone(Rank 19/22 in AIIDE 2015)Multi-Agent System, each controlled unit is an Agent.(Agent behave to avoid blocking situations, indecision, predictability.)Irons unis have simple behaviors like harass his main or keep alive

Terran

3rd RankLetaBot

Use A* and DFS algorithm to find a build place for block choke point (set up barricade)Use pathfinding algorithm for efficient resource gatheringUse Influence map to calculate units attack powerResource gathering algorithm based on linear algebra

Terran

7th Rank

protoss

Highlight

Acknowledgements

David Churchill (University of Alberta)Organizer of AIIDE StarCraft AI Competition

Michal Certicky and his team members (Czech Technical University in Prague)Organizer of Student StarCraft AI Tournament

QnA