660 AI agents tried mining a diamond in Minecraft and none could do it
A months-long contest challenging AI bots to mine diamonds in Minecraft ended after none of the 660 entrants could complete the task.
Contestants had four days to train their bot to perform the task, which involves forging an axe and digging beneath the earth to find a rare diamond in Minecraft's large, open ended virtual world.
"We challenge your agent to learn in only 8 million steps, a task humans can do with a single demonstration," the contest video reads.
The challenge, called MineRL, proved too much for any team. While AI bots have beat human players in ultra competitive and complex games like chess and Starcraft II, the training requirements and the shorter-than-usual preparation window for MineRL prevented anyone from winning.
Some bots dug themselves into the ground:
Neural networks playing @Minecraft, so exciting. :')— 𝚠𝚒𝚕𝚕𝚒𝚊𝚖 📀 (@wgussml) November 5, 2019
Here's a recent run of @Tviskaron's agent in the #MineRL comp, using option extraction to solve the Obtain Diamond task with hierarchical DQfD models.
w/ @rsalakhu @MSFTResearch & more!
Website: https://t.co/C504t4pRt8 pic.twitter.com/mOHdOG5umf
Microsoft Research, which helped organize the contest, said that it wasn't necessarily a surprise that no one could do it.
"The task we posed is very hard," Katja Hofmann, of Microsoft Research, told the BBC. "Finding a diamond in Minecraft takes many steps - from cutting trees, to making tools, to exploring caves and actually finding a diamond."
Researchers weren't allowed to train their bots for longer than four days. Researchers training AIs for other games, for example, train them to play at a pace that is equal to hundreds of years per day. The training cap for this contest, however, was set to see what developers could accomplish without a blank computing check.
Another hindering aspect of the competition involved researchers encouraging entrants to use imitation learning, a training method that teaches agents how to act by showing them examples. This contrasts with traditional reinforcement learning, which most startups use to train game playing bots.
If you want to see where human-machine competition is fiercest, look to games. Complex, rules-based games present the ideal environment to train AI systems, where the industry's top startups compete against the world's best players.