2018 StarCraft AI challenge CAS automation

  • Detail

On November, 2018, AAAI for interactive digital

entertainment was held at the University of Alberta. The results of the annual StarCraft AI challenge were announced at the meeting; This is also one of the most important competitions in the real-time strategy (RTs) game AI


a total of 27 teams participated in the 8th (2018) aiide StarCraft AI competition, including well-known institutions such as Samsung, the Institute of automation of the Chinese Academy of Sciences, Facebook, well-known universities such as Stanford University, and many independent teams. The challenge lasted about two weeks using 12 computers. After an average of about 2600 rounds for each BOT and about 100 rounds for each group of opponents, a total of 34694 rounds of wheel 1v1 competition, Samsung, Facebook and the Institute of automation of the Chinese Academy of Sciences won the top three with 95.91%, 90.86% and 87.11% respectively

aiide StarCraft AI challenge has been held for eight consecutive times since 2011. The competition takes StarCraft 1 as the carrier and aims to promote and evaluate the level of artificial intelligence used in real-time strategy games (RTs). In recent years, it has become the main platform and tool for game confrontation, reinforcement learning, imitation learning, multi task learning and even general AI algorithm research. Due to incomplete information including uncertain opponents and huge space for state and action, StarCraft AI is more challenging than go AI. Therefore, it has also attracted top research institutions at home and abroad, including deepmind, Facebook, Samsung, Stanford University and the Institute of automation of the Chinese Academy of Sciences. Relevant core technologies can be widely used in finance, economics, biology Political science, military and other fields

Introduction to the top three schemes

the third place is the Institute of automation of the Chinese Academy of Sciences, and the protoss bot "CSE"

in the 2017 StarCraft AI challenge, AI "CPAC" from the Institute of automation of the Chinese Academy of Sciences won the fourth place with a winning rate of 71%

in this year's competition, the protoss bot "CSE" from the intelligent systems and Engineering Research Center (crise) of the Institute of automation of the Chinese Academy of Sciences won the third place with 87.11%. It is understood that the center is an independent scientific research department of the Institute of automation, which is committed to the research on game confrontation and autonomous evolutionary intelligence, including perceptual intelligence and cognitive decision-making intelligence. The authors of CSE include Zhang Junge, Guo Wei, Yin Qiyue, Zhan Dong, Wang Qiwei, Hu Yihui, Shen Shengqi and Huang Kaiqi. In 2017, CPAC was also developed by key personnel of the team

"CSE" uses Protoss, which is driven by the combination of rules and learning. How to sneak attack with hidden knife absorbs the experience and knowledge of human players; When the construction queue is empty, "CSE" uses deep learning to predict the units to be built, and so on

second in Facebook, the Zerg bot "cherrypi"

compared with the 69% winning rate and the sixth place in last year's competition, the "Che further causes excess rrypi of industry capacity" from Facebook has been greatly improved this year, with a winning rate of 90.86%. "Cherrypi" first of all, there are 8 ~ 13 different strategies prepared in advance for the opponent of each race. Taking advantage of the characteristics of multiple 1v1 games, it will choose the best strategy according to the winning rate of previous games with the opponent. In addition, it also uses a pre trained machine model to estimate the winning rate of different available strategies according to the current competition state, and then switch to the strategy with the highest winning rate in some cases. This switching design also produces the effect of "hybrid strategy". The machine learning technology used in "cherrypi" also includes path finding search algorithms for avoiding obstacles, avoiding war and flying kites, learning building layout based on human data, offline reinforcement learning and learning, etc. Cherrypi is also the only BOT using GPU computing power this year

the first Samsung, the Terran bot "Saida"

we all know that StarCraft is particularly popular in South Korea. Since 2002, Korean professional starplayers have joined the professional team one after another and are sponsored by well-known companies such as Samsung and SK Telecom. Saida's success is likely to come from the help of its professional star players

according to the introduction, the core of the "Saida" BOT is to use a stable game strategy. It will first consider defense, and then wait for an opportunity to take away the opponent in the middle of the game. They believe that this strategy can deal with the strategies of most opponents and has the least weaknesses

"Saida" also applies some AI technologies. Based on the technology of ualbertabot (which will be mentioned later), they used a finite state machine to control units and buildings. Each unit and building has a specific state in each war situation. Multiple search algorithms are used to find enemy bases or areas where buildings can be built

in the development process, "Saida" also explored the use of CNN and codec structure to learn attack timing from human players, and the use of Multi-Agent Reinforcement Learning Method to learn unit micromanipulation in local games, but these technologies were not added to the version of this competition

detailed ranking

the official score form containing the scores of all teams and detailed competition data is shown in the figure below

other interesting matters in the competition

as the BOT with the highest winning rate in the competition, how does Saida perform against human players? The Samsung team tested amateur and professional players. "Saida" can beat amateur players and lose to professional players. However, they believe that the gap is not big, and the dawn of beating professional players is ahead

the organizer's knife Mending: now that there has been energy strike, a) according to the type of output source, there are electromechanical, hydraulic, pneumatic, electromagnetic and other important types; If the BOT is defeated by human beings, amateur players may never have a chance to win in the future...

most teams use a fixed race, and only one team can "random race" in 2016, 2017 and 2018. Obviously, a single race means giving up breadth, digging deeper into the effective strategies that have been found, and it is easier to achieve good results. The random race programs in 2017 and 2018 are from ualbertabot of Memorial University of Newfoundland in Canada. This year, they ranked 19th with a winning rate of 34.71%. Samsung Saida, which won the championship based on ualbertabot, obviously only borrowed technology to realize its own strategy

the strategy used by "Saida" obviously has great advantages. According to the winning rate fluctuation chart provided by the challenge organizer, "Saida" achieved the highest winning rate at the beginning of the competition and remained stable until the end. At the beginning of the second place in CSRI, the "second place" in CSRI began to fluctuate, and then the "second place" in CSRI soon stabilized

the blue line in the figure is "Saida" and the black line is "cherrypi". The experimental system should be firm and durable, and the green line is "CSE"

"cherrypi" GitHub address:

"locutus" GitHub address:


Copyright © 2011 JIN SHI