Cheong-mok Bae, Seonghoon Yoon, Kyung-Joong Kim
Cognition and Intelligence Lab (http://cilab.sejong.ac.kr)
Sejong University, Seoul, Republic of Korea
2017 IEEE CIG StarCraft AI Competition
(https://cilab.sejong.ac.kr/sc_competition/)
1
StarCraft (Blizzard, 1998)
• This real-time strategy (RTS) game has been known as
one of the most difficult video games to solve
• People think that this game might be the next target for AI
researchers after oriental Go success (Alpha Go).
• State-of-the-art AIs are still novice human level (big gap) (Kim 2016)
2
StarCraft II StarCraft RemasteredStarCraft Brood War
[1] M.-J. Kim, K.-J. Kim, S.-J. Kim, and A. K. Dey, “Evaluation of StarCraft Artificial Intelligence Competition Bots by
Experienced Human Players,” ACM CHI (Late-Breaking Work), 2016
History of StarCraft AI
• 2009 BWAPI (Brood War API) was first developed
• 2010 1st AIIDE StarCraft AI Competition
1st IEEE CIG StarCraft AI Competition
• 2011 1st Student StarCraft AI Competition (SSCAIT)
• 2010 ~ Now three annual AI Competitions (CIG/AIIDE/SSCAIT)
• 2017 Official StarCraft II API released
(Blizzard and Google DeepMind Team)
http://www.cs.mun.ca/~dchurchill/starcraftaicomp/history.shtml 3
Competition Setup
• StarCraft Brood War
• BWAPI (Brood War Programming Interface)
• C++ or JAVA or Proxy Bot (any programming language)
• Full-round robin style competition
• 5 maps used (randomly selected from official maps)
• Winner is the one with highest win ratio
• 125 rounds (190 games/round)
• Total 47,500 games (two weeks using 22 machines)
• Open-source policy
4
Number of Submissions
5
10 10
8
13
14
16
20
2011 2012 2013 2014 2015 2016 2017
20 Entrants from 13 Countries
New Entries Upgrade from 2016 Version
Re-entrance
of 2016 Version
# of
Entries
5 7 8
Bots
PurpleWave (USA)
McRave (Canada)
BigEyes (Korea)
Sling (Korea)
CasiaBot (China)
tscmoo (Norway,1st in 2016)
Iron (France, 2nd in 2016)
LetaBot (Netherland, 3rd in 2016)
ZZZKBot (Australia, 4th in 2016)
MegaBot (Brazil)
Tyr (Netherland)
Ziabot (Korea)
Aiur (France)
Bonjwa (Vietnam)
OpprimoBot (Sweden)
Salsa (Spain)
TerranUAB (Poland)
UAlbertaBot (Canada)
Overkill (China)
SRbotOne (Canada)
6
Race Distribution
Year Winner Race
2017 ? ?
2016 TSCMOO Terran
2015 ZZZBOT Zerg
2014 ICEBOT Terran
2013 SkyNet Protoss
2012 SkyNet Protoss
2011 SkyNet Protoss
7
Competition Maps
Hitchhiker1.1
TauCross1.1
Alchemist1.0
Andromeda1.0
Python1.3
Two Players Map
8
Three Players Map Four Players Map
Total Games Played
9
FILE I/O is allowed
AI Player can save
experience over rounds
to adapt strategy
AI Bots
10
PurpleWave (Protoss)
• Newcomer
• Author: Dan Gant (USA)
• https://github.com/dgant/PurpleWave
• Proxy bot written in Scala
• Decision making: Simple Hierarchical Task Network (e.g., “train
a probe”  “train probes continuously until saturation”)
• Micromangement: A hybrid squad/multi-agent approach
(Units have different goals but they collaborate when they’re
in the same squad)
11
CasiaBot (Zerg)
• Newcomer
• Author: Junliang Xing (China)
• Institute of Automation, Chinese Academy of Sciences
• https://sites.google.com/site/junliangxing/
• developed based on UAlbertaBot
• “deep learning models to automatically select the most
suitable strategy dynamically at the beginning of the game
and during its playing”
• supervised learning and reinforcement learning deep models
12
LetaBot (Terran)
• 3rd rank in 2016
• Author: Martin Rooijackers (Netherland)
• https://github.com/MartinRooijackers/LetaBot
• “The newest version of my bot uses build order from
liquidpedia. These build orders are what professional BW
players use as well.”
• “Text mining was used to acquire the build orders. If all goes
well my bot will be using MCTS this year.”
13
MegaBot (Protoss)
• 7th rank in 2016
• Author: Anderson Rocha Tavares (Brazil)
• https://github.com/andertavares/MegaBot
• “We use Skynet, Xelnaga and NUSBot as our portfolio.”
• “We use a Q-learning-like approach to update the values of
portfolio components. Also, we select the portfolio
component to activate via epsilon-greedy.”
• a similar approach in FTG-AI
( https://github.com/TiagoNO/MegaBot-FTG-AI )
14
Results Announcement
15
Result (1st ~ 3rd)
Rank Bot Name WinRace Lose Win RateGame
16
Congratulation!
Winner of 2017 IEEE CIG
StarCraft AI Competition
ZZZKBot by Chris Coxe
from Australia and Britain
1 ZZZKBot 1984 1628 356 82.06%
Famous for it’s very early attack strategy known as four
drone rush (4-pool build order). AI players have limited
ability to show solid opening defense.
2 tscmoo 1992 1541 451 77.36%
It’s slightly lower win ratio than the winner. It’s based on
many strategies on the three races (multi-agent approach).
It’s a newcomer with HTN and hybrid squad/multi-agent control
3 PurpleWave 2021 1360 661 67.29%
Last year
bot
Result (1st ~ 10th)
1 ZZZKBot 1984
2 tscmoo 1992
4 LetaBot 2026
5 UAlbertaBot 2005
6 Overkill 2024
7 CasiaBot 2003
8 Ziabot 2005
9 Iron 1988
10 Aiur 2086
3 PurpleWave 2021
Rank Bot Name WinRace Lose Win RateGame
1628
1541
1363
1315
1270
1246
1238
1225
1248
1360
356
451
663
690
754
757
767
763
838
661
82.06%
77.36%
67.28%
65.59%
62.75%
62.21%
61.75%
61.62%
59.83%
67.29%
In 2017, IEEE CIG provides
prizes to winners of game
competition.
Appointable candidates :
• Student
• Young Professional
• The bot should beat all
the sample controller.
Occupation
Public
Public
Public
Student
Last year
bot
Young
Professional
Last year
bot
Student
Young
Professional
Result (11th ~ 20th)
11 MegaBot 2013
12 McRave 1964
14 TerranUAB 1990
15 SRbotOne 1951
16 OpprimoBot 2059
17 Bonjwa 1943
18 Bigeyes 1976
19 Sling 1930
20 Salsa 2024
13 Tyr 2008
Rank Bot Name WinRace Lose Win RateGame
1105
941
732
666
632
596
557
349
94
890
838
908
1118
1285
1427
1347
1419
1581
1930
1023
54.89%
47.91%
36.78%
34.14%
30.69%
30.67%
28.19%
18.08%
4.64%
44.32%
Occupation
Young
Professional
Young
Professional
Last year
bot
Last year
bot
Last year
bot
Student
Last year
bot
Public
Last year
bot
Young
Professional
19
Future Issues
• Human vs. StarCraft AI
• We plan to organize a special event Human vs. StarCraft AI
• Unfortunately, the current level of AI Player is not competitive to
novice human player
• StarCraft II
• You can build StarCraft II AI using Blizzard’s official API and Google
DeepMind’s machine learning tool
• The API is not made by hacking and there is more chance to use
machine learning for real-time strategy game.
20
Dave Churchill
Organizer of AIIDE StarCraft AI Competition
Acknowledgements
Michal Certicky and his team
Organizer of Student StarCraft AI Tournament
21
Q & A
22

2017 CIG StarCraft

  • 1.
    Cheong-mok Bae, SeonghoonYoon, Kyung-Joong Kim Cognition and Intelligence Lab (http://cilab.sejong.ac.kr) Sejong University, Seoul, Republic of Korea 2017 IEEE CIG StarCraft AI Competition (https://cilab.sejong.ac.kr/sc_competition/) 1
  • 2.
    StarCraft (Blizzard, 1998) •This real-time strategy (RTS) game has been known as one of the most difficult video games to solve • People think that this game might be the next target for AI researchers after oriental Go success (Alpha Go). • State-of-the-art AIs are still novice human level (big gap) (Kim 2016) 2 StarCraft II StarCraft RemasteredStarCraft Brood War [1] M.-J. Kim, K.-J. Kim, S.-J. Kim, and A. K. Dey, “Evaluation of StarCraft Artificial Intelligence Competition Bots by Experienced Human Players,” ACM CHI (Late-Breaking Work), 2016
  • 3.
    History of StarCraftAI • 2009 BWAPI (Brood War API) was first developed • 2010 1st AIIDE StarCraft AI Competition 1st IEEE CIG StarCraft AI Competition • 2011 1st Student StarCraft AI Competition (SSCAIT) • 2010 ~ Now three annual AI Competitions (CIG/AIIDE/SSCAIT) • 2017 Official StarCraft II API released (Blizzard and Google DeepMind Team) http://www.cs.mun.ca/~dchurchill/starcraftaicomp/history.shtml 3
  • 4.
    Competition Setup • StarCraftBrood War • BWAPI (Brood War Programming Interface) • C++ or JAVA or Proxy Bot (any programming language) • Full-round robin style competition • 5 maps used (randomly selected from official maps) • Winner is the one with highest win ratio • 125 rounds (190 games/round) • Total 47,500 games (two weeks using 22 machines) • Open-source policy 4
  • 5.
    Number of Submissions 5 1010 8 13 14 16 20 2011 2012 2013 2014 2015 2016 2017
  • 6.
    20 Entrants from13 Countries New Entries Upgrade from 2016 Version Re-entrance of 2016 Version # of Entries 5 7 8 Bots PurpleWave (USA) McRave (Canada) BigEyes (Korea) Sling (Korea) CasiaBot (China) tscmoo (Norway,1st in 2016) Iron (France, 2nd in 2016) LetaBot (Netherland, 3rd in 2016) ZZZKBot (Australia, 4th in 2016) MegaBot (Brazil) Tyr (Netherland) Ziabot (Korea) Aiur (France) Bonjwa (Vietnam) OpprimoBot (Sweden) Salsa (Spain) TerranUAB (Poland) UAlbertaBot (Canada) Overkill (China) SRbotOne (Canada) 6
  • 7.
    Race Distribution Year WinnerRace 2017 ? ? 2016 TSCMOO Terran 2015 ZZZBOT Zerg 2014 ICEBOT Terran 2013 SkyNet Protoss 2012 SkyNet Protoss 2011 SkyNet Protoss 7
  • 8.
  • 9.
    Total Games Played 9 FILEI/O is allowed AI Player can save experience over rounds to adapt strategy
  • 10.
  • 11.
    PurpleWave (Protoss) • Newcomer •Author: Dan Gant (USA) • https://github.com/dgant/PurpleWave • Proxy bot written in Scala • Decision making: Simple Hierarchical Task Network (e.g., “train a probe”  “train probes continuously until saturation”) • Micromangement: A hybrid squad/multi-agent approach (Units have different goals but they collaborate when they’re in the same squad) 11
  • 12.
    CasiaBot (Zerg) • Newcomer •Author: Junliang Xing (China) • Institute of Automation, Chinese Academy of Sciences • https://sites.google.com/site/junliangxing/ • developed based on UAlbertaBot • “deep learning models to automatically select the most suitable strategy dynamically at the beginning of the game and during its playing” • supervised learning and reinforcement learning deep models 12
  • 13.
    LetaBot (Terran) • 3rdrank in 2016 • Author: Martin Rooijackers (Netherland) • https://github.com/MartinRooijackers/LetaBot • “The newest version of my bot uses build order from liquidpedia. These build orders are what professional BW players use as well.” • “Text mining was used to acquire the build orders. If all goes well my bot will be using MCTS this year.” 13
  • 14.
    MegaBot (Protoss) • 7thrank in 2016 • Author: Anderson Rocha Tavares (Brazil) • https://github.com/andertavares/MegaBot • “We use Skynet, Xelnaga and NUSBot as our portfolio.” • “We use a Q-learning-like approach to update the values of portfolio components. Also, we select the portfolio component to activate via epsilon-greedy.” • a similar approach in FTG-AI ( https://github.com/TiagoNO/MegaBot-FTG-AI ) 14
  • 15.
  • 16.
    Result (1st ~3rd) Rank Bot Name WinRace Lose Win RateGame 16 Congratulation! Winner of 2017 IEEE CIG StarCraft AI Competition ZZZKBot by Chris Coxe from Australia and Britain 1 ZZZKBot 1984 1628 356 82.06% Famous for it’s very early attack strategy known as four drone rush (4-pool build order). AI players have limited ability to show solid opening defense. 2 tscmoo 1992 1541 451 77.36% It’s slightly lower win ratio than the winner. It’s based on many strategies on the three races (multi-agent approach). It’s a newcomer with HTN and hybrid squad/multi-agent control 3 PurpleWave 2021 1360 661 67.29%
  • 17.
    Last year bot Result (1st~ 10th) 1 ZZZKBot 1984 2 tscmoo 1992 4 LetaBot 2026 5 UAlbertaBot 2005 6 Overkill 2024 7 CasiaBot 2003 8 Ziabot 2005 9 Iron 1988 10 Aiur 2086 3 PurpleWave 2021 Rank Bot Name WinRace Lose Win RateGame 1628 1541 1363 1315 1270 1246 1238 1225 1248 1360 356 451 663 690 754 757 767 763 838 661 82.06% 77.36% 67.28% 65.59% 62.75% 62.21% 61.75% 61.62% 59.83% 67.29% In 2017, IEEE CIG provides prizes to winners of game competition. Appointable candidates : • Student • Young Professional • The bot should beat all the sample controller. Occupation Public Public Public Student Last year bot Young Professional Last year bot Student Young Professional
  • 18.
    Result (11th ~20th) 11 MegaBot 2013 12 McRave 1964 14 TerranUAB 1990 15 SRbotOne 1951 16 OpprimoBot 2059 17 Bonjwa 1943 18 Bigeyes 1976 19 Sling 1930 20 Salsa 2024 13 Tyr 2008 Rank Bot Name WinRace Lose Win RateGame 1105 941 732 666 632 596 557 349 94 890 838 908 1118 1285 1427 1347 1419 1581 1930 1023 54.89% 47.91% 36.78% 34.14% 30.69% 30.67% 28.19% 18.08% 4.64% 44.32% Occupation Young Professional Young Professional Last year bot Last year bot Last year bot Student Last year bot Public Last year bot Young Professional
  • 19.
  • 20.
    Future Issues • Humanvs. StarCraft AI • We plan to organize a special event Human vs. StarCraft AI • Unfortunately, the current level of AI Player is not competitive to novice human player • StarCraft II • You can build StarCraft II AI using Blizzard’s official API and Google DeepMind’s machine learning tool • The API is not made by hacking and there is more chance to use machine learning for real-time strategy game. 20
  • 21.
    Dave Churchill Organizer ofAIIDE StarCraft AI Competition Acknowledgements Michal Certicky and his team Organizer of Student StarCraft AI Tournament 21
  • 22.

Editor's Notes

  • #18 The winner is ZZZKBot by Chris Coxe. The Bot’s win ratio is about 75%. This bot is famous for it’s very early attack strategy known as four drone rush. It’s still very effective to many other bots. It means that still AI players has limited ability to show solid opening defense. The 2nd ranker is TSCMOO with random race. It’s slightly lower win ratio than the winner. It’s based on many strategies on the three races. The 3rd ranker is the newcomer in this year, PurpleWave. Congratulation on the top three AI players.
  • #19 Thanks for all entries submitted in this year competition. Their bot source code and replays will be posted to our competition website in early Stember.