The Current State ofStarCraft AI Competitions and Bots

Michal CertickyDepartment of Computer Science

Czech Technical University in Praguecerticky@agents.fel.cvut.cz

David ChurchillDepartment of Computer Science

Memorial University of Newfoundlanddave.churchill@gmail.com

AbstractReal-Time Strategy (RTS) games have become anincreasingly popular test-bed for modern artificialintelligence techniques. With this rise in popular-ity has come the creation of several annual com-petitions, in which AI agents (bots) play the fullgame of StarCraft: Broodwar by Blizzard Enter-tainment. The three major annual StarCraft AICompetitions are the Student StarCraft AI Tourna-ment (SSCAIT), the Computational Intelligence inGames (CIG) competition, and the Artificial Intel-ligence and Interactive Digital Entertainment (AI-IDE) competition. In this paper we will give anoverview of the current state of these competitions,and the bots that compete in them.

IntroductionReal-time Strategy (RTS) games are a genre ofvideo games in which players manage economicand strategic tasks by gathering resources, buildingbases, increase their military power by researchingnew technologies and training units, and lead theminto battle against their opponent(s). They serveas an interesting domain for Artificial Intelligence(AI) research and education, since they representa well-defined, complex adversarial systems (Buro2004) which pose a number of interesting AI chal-lenges in the areas of planning, dealing with un-certainty, domain knowledge exploitation, task de-composition, spatial reasoning, and machine learn-ing (Ontanon et al. 2013).

Unlike turn-based abstract board games likechess and go, which can already be played by AI at

Copyright c© 2017, Association for the Advancementof Artificial Intelligence (www.aaai.org). All rights re-served.

super-human skill levels, RTS games are played inreal-time, meaning the state of the game will con-tinue to progress even if the player takes no ac-tion, and so actions must be decided in fractionsof a second. In addition to that, individual turnsin RTS games (game frames) can consist of issu-ing simultaneous actions to hundreds of units atany given time (Buro and Churchill 2012). This,together with their partially observable and non-deterministic nature, makes RTS game genre oneof the hardest game AI challenges today, attract-ing the attention of the academic research commu-nity, as well as commercial companies. For exam-ple, Facebook AI Research, Microsoft, and GoogleDeepMind have all recently expressed interest inusing the most popular RTS game of all time:Starcraft as a test environment for their AI research(Gibney 2016).

Meanwhile, the academic community has beenusing StarCraft as a domain for AI research sincethe advent of the Brood War Application Program-ming Interface (BWAPI) in 2009 (Heinermann2013). BWAPI allows programs to interact withthe game engine directly to play autonomouslyagainst human players or against other programs(bots). The introduction of BWAPI gave rise tonumerous scientific publications over last 8 years,dealing with all kinds of sub-problems inherentto RTS games. A comprehensive overview can befound in (Churchill et al. 2016), (Ontanon et al.2015) or (Ontanon et al. 2013).

In addition to AI research, StarCraft and BWAPIare often used for educational purposes as part ofAI-related courses at universities, including UCBerkeley (US), Washington State University (US),

University of Alberta (CA), Comenius Univer-sity (SK), Czech Technical University (CZ), Uni-versity of Zilina (SK) and most recently Techni-cal University Delft (NL), where a new courseentitled “Multi-agent systems in StarCraft” hasbeen opened for over 200 students. The educa-tional potential of StarCraft has recently been ex-tended even further, when Blizzard Entertainmentreleased the game entirely for free in April 2017.

Widespread use of StarCraft in research andeducation has lead to a creation of three annualStarCraft AI competitions existing until today. Thefirst competition was organized at the University ofCalifornia, Santa Cruz in 2010 as part of the AAAIArtificial Intelligence and Interactive Digital En-tertainment (AIIDE) conference program. The fol-lowing year gave rise to other two annual com-petitions – Student StarCraft AI Tournament (SS-CAIT), organized as a standalone long-term eventat Comenius University in Bratislava and CzechTechnical University in Prague, and CIG StarCraftAI competition collocated with IEEE Computa-tional Intelligence in Games (CIG) conference.

In this paper, we will talk about these threemajor StarCraft AI competitions and provide thelatest updates on each of them, with the follow-ing 3 sections detailing the SSCAIT, AIIDE, andCIG Starcraft AI Competitions. We will also takea closer look at the state of the research and brieflydescribe current participants (bots) and AI meth-ods they use.

SSCAIT: Student StarCraftAI Tournament

The Student StarCraft AI Tournament (SSCAIT) isthe StarCraft AI competition with the highest num-ber of total participants. It started as a part of anAI course at Comenius University, and initial sea-sons included several dozen student submissionsfrom this course, in addition to submissions fromacross the globe. Since then, SSCAIT started ac-cepting non-student participants and team submis-sions. There are three fundamental differences be-tween SSCAIT and the remaining two competi-tions:

1. SSCAIT is an online-only event. Unlike AIIDEor CIG, it is not co-located with a scientific con-ference or any other real-world event.

2. There are two phases of SSCAIT each year:

a competitive tournament phase, lasting for upto three weeks and a ladder phase which runsfor approximately eleven months each year. Inother words, SSCAIT is live at all times withonly a few short interruptions for maintenance.

3. Games are played one at a time and are publiclystreamed live on Twitch.tv1 and SmashCast.tv24 hours a day. The AIIDE and CIG competi-tions instead play as many games as possible atmaximum speed, with no public broadcast.

SSCAIT 2016-17 Updates & NewsThe activity of bot programmers and the generalpublic surrounding SSCAIT has grown consider-ably over last several months – mainly thanks to anumber of improvements of live stream and bettercommunity engagement during the Ladder phase,which was possible thanks to new members of theorganizing team.

First, the ladder phase was updated, with SS-CAIT introducing so-called “weekly reports”. Ev-ery weekend, there is a 1-2 hours long segmentof curated AI vs. AI matches with insightful com-mentary on the live stream.

Second, the voting system was implemented,allowing bot programmers and viewers to selectwhich bots will play the next ladder match on livestream (fig. 1). This not only supports viewer en-gagement, but also greatly simplifies bot debug-ging process. Bot programmers can now quicklytest their newest updates against specific oppo-nents. This change might have contributed to thesignificant increase in bot update frequency. Ap-proximately 5-6 bots are updated every day, in con-trast to 0-2 updates per week in 2015.

Another update was the introduction of “mini-tournaments” to SSCAIT. These are easily config-urable, irregular and unofficial short competitions,taking up to one day. The format of these minitour-naments and the selection of participants is usuallyup to the stream viewers and moderators.

Visual quality of the stream itself was improvedby perfecting the custom observer script (Matts-son, Vajda, and Certicky 2015) which now movesthe camera fluently to most interesting parts of thegame in real time and displays SSCAIT-related in-formation on top of the game. The stream was also

1http://www.twitch.tv/sscait

Figure 1: A web-based interface allowing SSCAITviewers and bot programmers to vote for the nextladder match.

Figure 2: SSCAIT live stream running in HD res-olution and controlled by the custom observerscript (Mattsson, Vajda, and Certicky 2015).

upgraded to run in HD resolution using a “resolu-tion hack” (fig. 2).

Furthermore, two additional metrics were addedto the ladder ranking system due to popular de-mand: ELO rating (Elo 1978), which is commonlyused in adversarial games like chess, and “SSCAITrank”, based on so-called “ICCUP ranking sys-tem”, typical for competitive StarCraft.

The overall number of stream views has in-creased to 376,920 views on Twitch.tv and ad-ditional 434,216 views on SmashCast.tv over thepast 12 months. The current number of active botson the ladder is 70.

Tournament Phase Updates

The 2016/17 installment of SSCAIT’s tournament

phase took place during three weeks at the endof December 2016 and beginning of January 2017and sported 45 participants. The tournament wasdivided into two divisions:

Student Division: Round Robin tournament of1980 games, where every bot played two gamesagainst every opponent. Only the bots created by asingle participant, who was a student at that time,were considered “student” bots and were eligiblefor victory in this division. Other bots were taggedas “mixed-division” bots (they played the games,but could not win the student division title). Win-ners of the student division in 2016/17 were:

1. LetaBot (Martin Rooijackers), University ofMaastricht (Netherlands) with 82 wins

2. Wulibot, University of Southern California(USA) with 63 wins

3. Zia Bot, Ulsan National Institute of Science andTechnology (South Korea) with 54 wins

The student division of SSCAIT exists so that thestudents stand a chance of winning in the presenceof more experienced, non-student participants andteam-created bots.

Mixed Division: After the student divisionended, sixteen bots with the most wins among allthe participants were selected for the additionalmixed division elimination bracket. Since two bots(Steamhammer and Zia bot) were tied for 16thplace, both with 54 wins, additional tie-breakermatch was scheduled to decide who gets 16thposition in the mixed division bracket.

The following 16 bots played in the mixeddivision elimination bracket: LetaBot, Steamham-mer, Overkill, WuliBot, ZZZbot, UAlbertaBot,IronBot, IceBot, Bereaver, BeeBot, XIMP, Skynet,Krasi0bot, Flash AI, KillerBot and tscmoo. Firstround (Ro16) matches consisted of only a singlegame, but Ro8 and Ro4 were played as “best of 3”matches. Finals were played as “best of 5”.

Interestingly, LetaBot (created by MartinRooijackers) managed to win the mixed divisionelimination bracket, in addition to winning thestudent division, after beating Krasi0bot (createdby Krasimir Krystev) 3 to 1 in the finals (fig. 3).LetaBot had hard-coded special strategies againstspecific opponents. It successfully executed an“SCV rush” in the final games – a strategy for

Figure 3: SSCAIT 2016/17 mixed division elimi-nation bracket.

which Krasi0bot was not prepared.

All the elimination bracket games were pub-lished as videos with commentary on SSCAITYouTube channel2 and as replay files on SSCAITwebsite.3 Tournament winner, Martin Rooijackers,has recently been invited to have a talk about thisachievement at a business-oriented AI conferenceWorld Summit AI.4

AIIDE: Artificial Intelligence andInteractive Digital Entertainment

The AIIDE Starcraft AI Competition is the longestrunning annual Starcraft competition, and has beenheld every year since 2010 along with the AAAIArtificial Intelligence and Interactive Digital En-tertainment conference. Unlike the CIG and SS-CAIT competitions, the AIIDE competition re-quires that all bots be open source, and that theirsource code will be published for public down-load after the competition has finished. Running24 hours a day for 2 weeks with games playedat super-human speed, the competition is a sin-gle round-robin format with the winner being thebot with the highest win percentage when the timelimit has been reached.

2http://youtube.com/playlist?list=PLS6Qj916df6K9z3dOLQhCS0KatlvBqhbE

3http://sscaitournament.com/index.php?action=2016

4http://worldsummit.ai/

AIIDE 2016-17 Updates & NewsThe 2016 AIIDE competition had a total of 21competitors, and the round robin games ran on 12machines for nearly two weeks. A total of 90 roundrobin rounds were completed, with each bot play-ing 1800 games. No new rules or maps were usedfor the 2016 tournament that were not in placefor the 2015 tournament. As the AIIDE competi-tion was held shortly after the CIG competition,many of the submissions were the same, which isreflected in the results of both competitions. Thetop four finishers can be seen in Figure 4, withIron placing 1st, ZZZKbot placing 2nd, and tsc-moo coming in 3rd place.

Figure 4: Results of the top 4 finishers in the 2016AIIDE competition. Iron and LetaBot are Terranbots, while ZZZBot and tscmoo played as Zerg.

In 2016, the AIIDE competition website was up-dated to include an archive5 of data and results ofall of the annual AIIDE and CIG competitions, in-cluding final results, bot source code and binarydownload links, and information about each bot.

The 2017 AIIDE competition will have severalupdates:

• An updated map pool consisting of 10 newmaps.

• Updated tournament managing software capa-ble of playing more games in the same periodof time.

• Support for BWAPI version 4.2.0

• Bots that achieved a 30% or higher win rate inthe 2016 competition will be carried forward to2017

• Plans to support GPU computation for bots

5http://www.cs.mun.ca/˜dchurchill/starcraftaicomp/archive.shtml

CIG: ComputationalIntelligence in Games

The CIG StarCraft RTS AI Competition is a partof the program of the IEEE Computational In-telligence in Games (CIG) conference since Au-gust 2010. Similarly to AIIDE competition, all thebot games are run separately on several comput-ers prior to the conference and the results are thenannounced during the event. Unlike in AIIDE andSSCAIT, the map pool of the CIG competition isnot known in advance for the competitors. Also,the CIG competition no longer enforces an opensource code requirement to compete.

CIG 2016-17 Updates & NewsCIG 2017 is scheduled to happen in New YorkCity, USA on late August 2017, which is just afterthe submission deadline of this paper. Therefore,we will report on CIG 2016, which happened onSeptember 2016.

The 2016 installment of CIG competition hosted16 participants, out of which 9 were new orupdated bots and 7 were re-entries from pre-vious year. CIG 2016 competition was dividedinto two stages: The Qualifying stage and the Fi-nal stage. The first qualifying stage consisted ofRound Robin tournament between all 16 partici-pating bots. In total, 11988 games were played inthis stage. After the qualifying stage, best 8 botswere selected to proceed to the final stage: tscmoo,IronBot, LetaBot, ZZZbot, Overkill, UAlbertaBot,MegaBot and Aiur.

All the persistent files accumulated by the botsin the qualifying stage were deleted before enter-ing the finals. This stage consisted of 2799 RoundRobin games between the 8 bots. All the games ranon 17 computers for 8 days. The winner of the fi-nal stage, tscmoo bot created by Vegard Mella fromNorway, was announced the overall winner of CIG2016 tournament with 456 wins and 65.14% winrate in the final stage. Detailed results are depictedin Figure 5.

Current StarCraft BotsOver the years, StarCraft AI competitions havemotivated many individuals and groups to combineand integrate various AI techniques and methodsinto complete bots, capable of playing completeStarCraft 1v1 games.

Figure 5: Detailed results of the CIG 2016 compe-tition final stage.

In this section, we provide an overview of a se-lection of bots (in alphabetical order) and discusssome of the AI approaches they implement. Weonly mention those bots that are currently activein one of the competitions, have recently been up-dated, and employ some more complex AI tech-niques (we do not mention simple hard-coded orrule-based bots).

• AIlien: AIlien is a relatively new Zerg bot. Vari-ous kinds of its decisions rely on “scoring” sys-tems. Other than that, it employs state machines– especially for higher-level strategy and macrodecisions.

• GarmBot: GarmBot is organized as a multi-agent system using the blackboard architecture.Every unit is controlled by a single agent, im-plemented as a state machine.

• Ian Nicholas DaCosta: This Protoss bot uses ge-netic algorithms for targeting, and detecting en-emy army threat levels. Supervised learning wasapplied for the detection of opponent’s strategiesand builds.

• KaonBot: For resource and unit allocation basedon prioritized needs, KaonBot applies a compet-ing priority algorithm. According to the author,the bot will learn these priorities from experi-ence in future releases.

• Krasi0bot: Krasi0bot has been around for manyyears, but it is still being actively developed.Even though it originally started as a rule-basedbot, it currently makes some use of genetic al-gorithms, neural networks and potential fields.The author also actively experiments with vari-ous other techniques at the moment.

• LetaBot: The most interesting technique usedby Martin Rooijackers’ LetaBot is Monte Carlo

2010 2011 2012 2013 2014 2015 2016

AIIDE 17 13 10 8 18 22 21

CIG 10 10 8 13 14 18

SSCAIT 50 52 50 42 46 46

0

10

20

30

40

50

60

Starcraft AI Competitions - Total Entrants

2010 2011 2012 2013 2014 2015 2016

AIIDE 70 2340 8279 5579 10251 20788 18882

CIG 40 4050 2500 4680 2730 14787

SSCAIT 100 1326 1190 861 1035 1980

0

5000

10000

15000

20000

25000

Starcraft AI Competitions - Total Games Played

Figure 6: Statistics for each of the 3 major annual StarCraft AI Competitions: AIIDE, CIG, and SSCAIT,since the first competition in 2010. Shown on the left is the number of total entrants for each competition,and on the right are the total number of games played in each competition.

Tree Search (MCTS) algorithm, which is used toplan the movement of squads (groups of units)around the map. A similar approach has pre-viously been used by the author of Nova bot,Alberto Uriarte (Uriarte and Ontanon 2014).An implementation MCTS algorithm for squadmovement has also been very recently releasedin form of ready-to-use library StarAlgo.6 In ad-dition to MCTS, LetaBot employs cooperativepathfinding for resource gathering and text min-ing to extract build orders directly from Liqui-pedia articles.7

• MegaBot: For every game, MegaBot (Tavareset al. 2016) chooses one of three approaches,each of which is implemented as a different bot(Skynet, Xelnaga or NUSBot). Algorithm selec-tion is modeled as a multi-armed bandit. At thebeginning of the game, an algorithm is selectedusing epsilon-greedy strategy. After the game,the reward is perceived (+1, 0, -1 for victory,draw and loss, respectively) and the value of theselected algorithm is updated via an incremen-tal version of recency-weighted exponential av-erage (Q-learning update rule).

• Monica / Maria / Brenda: Zerg, Protoss andTerran bots employing a game simulation in-side the BEAM Erlang/OTP VM. It usesTorchCraft (Synnaeve et al. 2016) – a library formachine learning research on RTS games. Theunit logic is written in Lua.

6http://github.com/vnikk/StarAlgo7http://wiki.teamliquid.net/

starcraft

• PurpleWave: The decision making of Purple-Wave bot is mainly based on hierarchical tasknetworks. For the micromanagement, it uses ahybrid squad/multi-agent approach and nearestneighbors clustering. The bot then simulates theoutcomes of battles and suggests tactics for thesquads by min-maxing tactical approaches byeach side (e.g. “charge in”, “run away”, or “fightwith workers”). In the end, each unit takes thetactical suggestion under advisement, but be-haves independently. The units choose betweenaproximately two dozen simple, reusable state-less behaviors. Uses heuristics including poten-tial fields for the movement.

• StarcraftGP: StarcraftGP is the first StarCraftmeta-bot – a program that autonomouslycreates a program that autonomously playsStarCraft (Garcıa-Sanchez et al. 2015). Cur-rently, StarcraftGP v0.1 is using (Linear) Ge-netic Programming and it is able to directlywrite C++ code. Its first creations, namely Salsaand Tequila, have been the first bots not writtenby a human to participate in international com-petitions.

• Steamhammer / Randomhammer: Zerg botSteamhammer and its random-race version Ran-domhammer both employ sophisticated combatsimulation with alpha-beta search and portfoliosearch to predict the outcome of battles. Thebots also use hierarchical reactive control forthe units. For Protoss and Terran production,Randomhammer uses branch-and-bound search,while Zerg production is currently rule-based.

• tscmoo: tscmoo uses no external libraries: it hasits own combat simulation code to predict theoutcome of battles (while others typically usethe SparCraft combat simulation package8), itdoes not use BWTA9 to analyze the terrain andit even has its own threat-aware pathfinding forindividual units. The bot can use many differentstrategies and selects among them based on theirsuccess in previous games. Recent versions ofthe bot experimented with recurrent neural net-works for high-level strategy and build order de-cisions.

• Vaclav Bayer: Q-learning / ReinforcementLearning and Markov Decision Processes areused by this bot – mainly to select the best buildorder with respect to opponent’s strategy.

• Zia Bot: Despite the fact that most of Zia Bot’sfunctionality is heuristic and rule-based, the bottries to recognize and remember all the strate-gies used by its opponents and by itself. Thismemory is used to select the better performinggame plans in the following games.

ConclusionIn this paper we have given an overview of the 3major annual StarCraft AI competitions, as wellas some of the top performing bots that competein them. As seen in Figure 6, each year, participa-tion in these competitions has continued to rise, aswell as the number of games played between botsin the competitions. In the past 2-3 years, the botsin these competitions have become more strategi-cally complex and functionally robust, employinga range of state-of-the-art AI techniques from thefields of heuristic search, machine learning, neuralnetworks, and reinforcement learning. While thebots currently play at an amateur human level, wehope that advancing RTS AI techniques coupledwith the recent involvement of industry research,they will be able to compete with human expertsin the next few years.

ReferencesBuro, M., and Churchill, D. 2012. Real-time strat-egy game competitions. AI Magazine 33(3):106.

8http://github.com/davechurchill/ualbertabot/wiki/SparCraft-Home

9http://bitbucket.org/auriarte/bwta2

Buro, M. 2004. Call for AI research in RTS games.In Proceedings of the 4th Workshop on Challengesin Game AI, 139–142.Churchill, D.; Preuss, M.; Richoux, F.; Synnaeve,G.; Uriarte, A.; Ontanon, S.; and Certicky, M.2016. Starcraft bots and competitions.Elo, A. E. 1978. The rating of chessplayers, pastand present. Arco Pub.Garcıa-Sanchez, P.; Tonda, A.; Mora, A. M.;Squillero, G.; and Merelo, J. 2015. Towards au-tomatic starcraft strategy generation using geneticprogramming. In Computational Intelligence andGames (CIG), 2015 IEEE Conference on, 284–291. IEEE.Gibney, E. 2016. What google’s winning go algo-rithm will do next. Nature 531(7594):284–285.Heinermann, A. 2013. Broodwar API. https://github.com/bwapi/bwapi.Mattsson, B. P.; Vajda, T.; and Certicky, M. 2015.Automatic observer script for StarCraft: BroodWar bot games (technical report). arXiv preprintarXiv:1505.00278.Ontanon, S.; Synnaeve, G.; Uriarte, A.; Richoux,F.; Churchill, D.; and Preuss, M. 2013. A surveyof real-time strategy game AI research and compe-tition in StarCraft. IEEE Transactions on Compu-tational Intelligence and AI in Games 5:293–311.Ontanon, S.; Synnaeve, G.; Uriarte, A.; Richoux,F.; Churchill, D.; and Preuss, M. 2015. RTS AI:Problems and Techniques.Synnaeve, G.; Nardelli, N.; Auvolat, A.; Chintala,S.; Lacroix, T.; Lin, Z.; Richoux, F.; and Usunier,N. 2016. Torchcraft: a library for machine learn-ing research on real-time strategy games. arXivpreprint arXiv:1611.00625.Tavares, A.; Azpurua, H.; Santos, A.; andChaimowicz, L. 2016. Rock, paper, starcraft:Strategy selection in real-time strategy games. In12th Artificial Intelligence and Interactive DigitalEntertainment Conference (AIIDE), 93–99.Uriarte, A., and Ontanon, S. 2014. High-level rep-resentations for game-tree search in rts games. InTenth Artificial Intelligence and Interactive DigitalEntertainment Conference.