Notes

Minecraft Diamond Challenge Leaves AI Creators Stuck

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



12 December 2019



It takes minutes for most new Minecraft players to figure out how to find the diamonds that are essential to the game, but training artificial intelligence to do it is more difficult than we expected.



Over the summer, Minecraft publisher Microsoft and other organizations challenged developers to design AI agents that could locate the most sought-after gems.



Most can crack it in their first session.



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



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



The objective was to determine whether the issue could be solved without requiring massive amounts of computing power.
mcnames


Despite not having a winner, one organiser said she was "hugely amazed by" some of the participants.



"The task we set out to solve is very hard," said Katja Hofmann, a principal researcher at Microsoft Research. "Finding a Minecraft diamond is a process that involves many steps, from cutting down trees to creating tools, from exploring caves to actually finding a diamond.



"While no agent submitted has yet solved the problem completely but they have made a number of steps forward and developed some of the tools required in the process."



Mining diamonds



Since its debut in the year 2011, Minecraft has been extremely popular.



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



Diamond is one of the most important resources in Minecraft as it can be used to make powerful armour and powerful weapons.



However, to obtain the precious stone, a player must first go through a series of other steps.



Jules Portelly, a Minecraft player, said that it shouldn't take more then 20 minutes to collect your first diamonds if you're comfortable with the game.



The entry requirements were that the entrants only be allowed to make use of one graphics processing unit (GPU), and they must complete four days of training. AI systems often require months or even years of gaming time to master games such as StarCraft II.



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



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



"So they had to learn the concept about making tools, locating resources and then finding a gem."



Minecraft to run experiments with artificial intelligence



DeepMind AI is awarded Grandmaster status at Starcraft 2



AI assistant to help Minecraft players



The developers were asked to develop programs that learn by imitation, using a technique called "imitation learning".



This is where you attempt to make AI agents to take the best approach to solving a task by making them imitate humans or other software.



It differs from relying only on "reinforcement learning", in which an agent is effectively trained to determine the most effective solution via a process of trial and error, without using the knowledge of previous experiences.



AI for all



Researchers have found that using reinforcement learning by itself can sometimes deliver superior results.



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



However, this "pure" method usually requires a lot more computing power, which makes it prohibitively expensive for researchers other than large organizations or government agencies.



William Guss, a Carnegie Mellon University PhD student, was the main organizer of the contest. He said that the purpose of the competition was to show that "throwing huge compute at problems isn't always the best way to advance the state of the art as a field."



He said, "It works against democratising access to these systems of reinforcement learning and leaves the capacity to train agents in complex environments to corporations that have swathes to compute."



However, the outcome could be a good indicator of the advantages these organizations with good funding have.



Minecraft graphics to be more realistic'



19 August 2019



Minecraft player loses five-year-long game



29 April 2019


Homepage: https://mcnames.net/