DeepMind Releases Tools to Train Artificial Intelligence on StarCraft II

DeepMind and Blizzard have released tools to let AI researchers create AIs that can fairly compete against human players in the game of StarCraft II. The bots will only be able to see and do the same things human players can. They will not have the unfair advantage as previous bots have, such as superhuman click speed.

Experts say that for an AI to fairly beat a human at StarCraft II there will likely need to be numerous breakthroughs, current machine-learning algorithms won’t be enough.

But when AI does win at StarCraft II, it will be a far greater achievement than AlphaGo’s victories over the best human players of Go.

DeepMind and Blizzard also are making public a the data from 65,000 past StarCraft II games for reserachers to use in the training of their AIs and they say the archive will grow by around 500,000 games a month.

Previous StarCraft II API layer by DeepMind:

The original StarCraft is also widely used by AI and ML researchers, who compete annually in the AIIDE bot competition. StarCraft’s usefulness among researchers is stems from its complex, multi-layered gameplay.

“For example, while the objective of the game is to beat the opponent, the player must also carry out and balance a number of sub-goals, such as gathering resources or building structures. In addition, a game can take from a few minutes to one hour to complete, meaning actions taken early in the game may not pay-off for a long time. Finally, the map is only partially observed, meaning agents must use a combination of memory and planning to succeed.” Said DeepMind in its release statement.

Link to DeepMind’s Github release.

In most previous games that AI has vanquished humans in such as chess and Go, all of your opponent’s moves and pieces are viewable, which makes them perfect information games and much easier for AIs to train on.

StarCraft II is an imperfect information game, your enemies troops or and buildings can be hidden by the fog of war, which forces you to to plan an act based on your limited view to predict what your enemy is planning.

The complexity of StarCraft II is also far beyond any game AI as conquered before. The number of possible board configurations in the game of Go is greater than the number of atoms in the universe, 1 followed by 170 zeros, researchers estimate that you would need to add at least 100 more zeros to get the complexity of StarCraft II.

Researchers predict it will be at least five years before a StarCraft II AI can fairly beat a human. But many experts predicted even longer time frames for Go—right before AlphaGo defeated Lee Sedol.

Facebook Comments