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.


This content now needs to be approved by community moderators.


This essay was submitted and is waiting for review.

When will an AI achieve competency in the Atari classic Montezuma’s Revenge?


Reinforcement learning is a type of machine learning which focuses on methods that enable agents to learn to maximize some posited conception of cumulative reward. It has been become a core method of AI and machine learning research and practice.

The Arcade Learning Environment (ALE) is a platform that allows AI researchers to develop and evaluate algorithms across a wide array of Atari 2600 games in hopes of helping to spawn more general and domain-independent AI technology. The ALE's Atari games have been used for testing reinforcement learning algorithms in AI research since researchers at DeepMind Technologies applied the first deep learning model in 2013 to learn control policies directly from sensory input — namely, using a convolutional neural network. In their model, the input was raw pixels and output was a value function estimating future rewards.

In 2015, the then-acquired-by-Google DeepMind used sensible pseudo-counts from raw pixels and transformed those pseudo-counts into "intrinsic rewards" to learn to play a number of Atari 2600 games. In particular, for Montezuma's Revenge (the reputed most difficult Atari 2600 game), the AI with "intrinsic rewards" was able to explore 15 out of the 24 rooms on the first level out of three. The same AI without "intrinsic rewards" only explored 2 out of 24.

In Montezuma's Revenge, an AI can show off its ability to explore its environment by climbing down ladders, and then jumping skeletons in order to retrieve keys — demonstrating long-term planning ability and so-called "artificial curiosity".

We ask:

When will an AI be able to explore all the rooms on the first level of Montezuma’s revenge in less than or equal to 50 million frames of training?

As of question launch (July 2017), state of the art is 15 out of 24 rooms explored in 50 million frames of training by Google's DeepMind in 2015; see paper & video.

Resolution will occur when a credible paper or video is produced of an AI agent exploring all 24 trap-filled rooms of Montezuma's Revenge in less than or equal to 50 million frames of training (without previously being exposed to Montezuma's revenge or an essentially similar game, or using training data or code based on example solutions).

Make a 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.