It seems that teaching artificial intelligence (AI) to play Minecraft is a big challenge, according tomedia. The MineRL competition, announced earlier this year, requires the research team to develop AnI robots that can successfully mine diamonds in Minecraft. This is not an impossible task, but it does require the basics of the game.
AI players need to know how to cut trees, craft crossbows, explore underground caves and avoid monsters and lava. Most adults with these skills learn a few hours into the game, or faster by watching a youTube tutorial.
But according to results to be presented at the AI Conference NeurIPS, none of the 660 contestants in the MineRL competition will be able to complete the challenge. Although robots can learn intermediate steps, such as building a furnace to synthesize a crossbow, no AI robot can successfully find diamonds.
“The task we’re proposing is very difficult,” Katja Hofmann, principal researcher at Microsoft Research, who is involved in organizing the challenge, told the media. “
This may come as a surprise, especially when it is thought that AI has overtaken humans in games such as chess, Go and Dota 2. But it reflects the important limitations of the technology and the limitations that MineRL referees have set for real challenge teams.
MineRL robots are required to learn using a combination of imitation learning and intensive learning. Although AI typically takes on big challenges by combining these two methods, like the AlphaGo system, which starts by entering data from old games to learn to play chess, and then by playing over and over again to hone its skills and ultimately surpass ingenuity, However, MineRL robots that take a similar approach have relatively limited resources available.
It’s like the difference between a big league team, which has access to a wide variety of resources — coaches, nutritionists, the best equipment money can buy — and a small team, while the latter has very limited resources. MineRL’s approach may seem unfair, but it actually reflects the challenge of integrating AI into the real world, as few companies and research labs can compete with Google’s DeepMind.