Minecraft Diamond Challenge Leaves AI Creators Stumped

Minecraft diamond challenge leaves AI creators stumped By Sam Shead Technology reporter

12 December 2019

It takes just a few minutes for new Minecraft players to work out how to find the diamonds that are crucial to the game, but educating artificial intelligence to do it has proved harder than expected. Minecraft servers

In the summer, Minecraft publisher Microsoft and other organisations challenged coders to develop AI agents that could identify the sought-after gems.

Most people can master it in their first session.

However none of the more than 660 entries submitted met the challenge.

The official announcement of the results of the MineRL competition which is pronounced "mineral" is expected to be announced at the NeurIPSAI conference in Vancouver, Canada on Saturday.

The objective was to determine if the problem could be solved with a minimum of computing power.

Despite the lack of any winners One of the organisers said she was still "hugely impressed" by the participants.

"The task we asked for is very hard," said Katja Hofmann who is a principal researcher with Microsoft Research. "Finding a diamond in Minecraft requires a variety of steps ranging from cutting down trees, making tools, to exploring caves and actually finding a diamond.

"While the submitted agents aren't capable of solving the problem fully, they have made significant progress and have learned how to utilize many of the tools needed."

Mining diamonds

Since its launch in 2011, Minecraft has been wildly popular.

More than 180 million copies of the game that is open-world have been sold, and the game has more than 112 million monthly active players.

Diamond is among the most important resources in Minecraft because it can be used to make strong armor and powerful weapons.

To obtain the precious gemstone However, the player must complete several other steps.

"If you're already familiar with the game it shouldn't take you more than 20 minutes to earn your first diamonds," Minecraft player Jules Portelly told the BBC.

Participants were limited to use a single graphics processing unit (GPU) and four days of training. To put it in perspective, AI systems usually need months or years of game play to master titles like StarCraft II.

To help them train their systems, a small Minecraft dataset of 60 million frames of player data was made available to participants.

"At the start of every episode they spawned in a procedurally-generated Minecraft world," explained Dr Hofmann.

"So they really needed to learn the concept of finding resources creating tools and locating a diamond."

Minecraft for artificial intelligence research

DeepMind AI reaches Grandmaster status at Starcraft 2

AI assistant to assist Minecraft players

The organizers wanted coders to develop programs that could be learned through imitation, a method known as "imitation learning".

This means trying to convince AI agents to adopt the most effective approach by making them mimic the way humans or other software can do to complete a task.

It contrasts with relying solely on "reinforcement learning", in which agents are trained to determine the most effective solution through trial and error, not using the knowledge of previous experiences.

All AI

Research has demonstrated that reinforcement learning can sometimes result in better results.

For instance, DeepMind's AlphaGo Zero program trumped one of the research hub's earlier initiatives, which employed reinforcement learning and the study of labelled data from human play to learn the game of Go.

This "pure" method requires more computing power and is too expensive for researchers not working in large companies or governments.

William Guss, the main competition organiser and a PhD student at Carnegie Mellon University, told the BBC that the point of the competition was to demonstrate that "throwing massive computing at problems isn't always the right way for us to push the limits of the art in the field".

He added: "It works directly against increasing access to these systems of reinforcement learning and leaves the capacity to train agents in difficult environments to corporations with swathes of compute."

However, the final outcome could reveal the advantages these entities with a good financial position have.

Minecraft graphics to be more 'realistic'

19 August 2019

A Minecraft player is unable to play for five years.

29 April 2019