DeepMind, a developer of Go AI "AlphaGo" and "DQN", is challenging to build the strongest AI with "Starcraft 2"

When will Artificial Intelligence (AI) overtake humans in which field?In the research that says "Starcraft will be overtaken by AI from around 2018 to around 2033", actually it is "Starcraft 2Deep Mind and Blizzard are starting to develop AI which can play better than a human being.

DeepMind and Blizzard open StarCraft II as an AI research environment | DeepMind
https://deepmind.com/blog/deepmind-and-blizzard-open-starcraft-ii-ai-research-environment/

Blizzard and Deep Mind turn StarCraft II into an AI research lab | TechCrunch
https://techcrunch.com/2017/08/09/blizzard-and-deepmind-turn-starcraft-ii-into-an-ai-research-lab/

Google's parent company, Alphabet's AI research companyDeep Mindis. Speaking of Deep Mind, Go AI who defeated the strongest Chinese players in China "AlphaGo"AI to learn games themselves and play more than humans"DQNIt is a company that has developed such as.

Such DeepMind is a game development company "Blizzard" held in 2016Blizzcon 2016"We announced a plan to collaborate with SF strategy game" Starcraft 2 "to build an open AI research environment.

Google 's Artificial Intelligence Team "Deep Mind" who defeated the Go champion next challenges the game "Starcraft 2" - GIGAZINE


This plan has actually started, and DeepMind and Blizzard offer the latest tools including machine learning API, large game replay dataset, open source DeepMind toolset, and more.

The following movies are one of the tools provided for AI research "PySC 2"The action of the movie. PySC 2 is a tool that provides a flexible and easy-to-use interface for AI to play games, while maintaining the basic visual and spatial elements of the game, elements such as unit type, state, It is displayed in "Feature layer" separated from each other.

DeepMind open source PySC 2 toolset for Starcraft II - YouTube


In addition, Blizzard is the latest Starcraft 2 API, offering packages for Linux, Windows and Macs designed to run on the cloud. In addition, it seems that AI is steadily increasing data for self learning by allowing AI vs. AI battle offline and having a human player provide anonymous game replay from a human player is.

Starcraft 2 basically seems to be a very useful environment for AI research, and this is influenced by the complexity and diversity of the game, and the fact that there are multiple paths of victory in individual games about. In addition, players need to advance various factors simultaneously such as management and generation of resources, command of military, construction of defensive facilities, etc. In addition, since not all information on games can be used immediately, players Since it is necessary to take actions based on assumptions and predictions, DeepMind's AI will require much more complicated tasks than those challenged so far.

However, it is too difficult to develop an AI that will handle all at the same time, so it seems that we are decomposing tasks and conducting research. By collecting resources and constructing facilities, we break down the actions in the game to make them executable and finally create AI that can win against human players.


The goal of DeepMind is to create an AI that plays StarCraft 2 better than a human player. DeepMind hopes that this will be useful for promoting existing research, and for that reason we are releasing research communities and tools open source.