Your submission is now in Draft mode.

Once it's ready, please submit your draft for review by our team of Community Moderators. Thank you!

Submit Essay

Once you submit your essay, you can no longer edit it.

Pending

This content now needs to be approved by community moderators.

Submitted

This essay was submitted and is waiting for review.

When will an RL agent match the best human performance on Starcraft II with no domain-specific hardcoded knowledge, trained using no more than $10,000 of compute of public hardware?

Question

Starcraft II is a hugely popular online, real-time strategy game, which provides a good testing ground for AI systems. As DeepMind wrote when announcing their Starcraft II project:

"[...] 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."

Note that this question does not just ask when Starcraft can be solved with any amount of effort, but rather relativises the performance to a publicly verifiable benchmark. We shall say that the game is solved when the agent is at least as good as the best human.


Resolution date will be set to the earliest of the following dates:

  • Publication date of a credible paper, blog-post, video or similar demonstrating an AI achieving the feat in question

  • A date earlier than the publication date, but referenced in a credible paper, blog-post, video or similar, by which the feat was achieved (similar to how DeepMind kept AlphaGo's victory over European champion Fan Hui secret from October 2015 to January 2016, in order to coincide with the publication of the corresponding Nature paper)

  • A date such that an expert council of technical AI researchers agree (by majority vote) that it is >=95% likely that the feat could have been carried out by that date. This means not just the date when the computational resources and algorithmic insights were available, but the date were they could have been fully deployed to solve the problem. For example, think the end and not the beginning of the AlphaGo project. [1] [2]

The $10,000 are 2018 US dollars.


[1] The point of this counterfactual resolution condition is as follows. Not all trajectories to advanced AI pass by a Starcraft victory. For example, it is possible that OpenAI were to impressively solve DOTA in full generality, and DeepMind subsequently drop their Starcraft project. In such a world, there might be a point at which it is clear that the feat in the question is achievable, even though no one has actually bothered to implement the experiment.
This is similar to how the release of the AlphaZero agent (playing chess, Go and shogi) should give us >=95% confidence that it is possible to play Othello at superhuman level without domain-specific knowledge, even though no one (to the author's knowledge) ran that experiment.
[2] More details regarding the council can be found here.

Make a Prediction

Prediction

Note: this question resolved before its original close time. All of your predictions came after the resolution, so you did not gain (or lose) any points for it.

Note: this question resolved before its original close time. You earned points up until the question resolution, but not afterwards.

Current points depend on your prediction, the community's prediction, and the result. Your total earned points are averaged over the lifetime of the question, so predict early to get as many points as possible! See the FAQ.