Microsoft has shown that Minecraft can be an excellent test-bed for innovative AI development as its Minecraft AI challenge stumps machine learning systems.
Artificial intelligence and machine learning have managed to solve some pretty complex problems in recent years, but it struggles with some problems that humans find easy to solve. Google's DeepMind AI researchers recently managed to train an AI to play the video game Starcraft II to such a high standard that it could beat most top players, but it couldn't cope with many unexpected and improvisational strategies.
Minecraft has seen popular use as a learning tool in classrooms to engage young students in everything from physics to electronics, and now Microsoft is using it as a test-bed for artificial intelligence. The Minecraft AI challenge tasked coders around the world with building AI agents that could achieve several simple objectives that players tend to work out quickly: hunt for Diamonds.
Diamonds in Minecraft are found deep underground and normally near lava, so to get there requires a number of serious challenges for an AI to master: Navigating the game world, modify it by mining, craft tools, survive environmental threats and enemies, and finally locate and obtain the diamonds. Players familiar with the game should be able to do this very quickly, but it's a tough problem for an AI to crack.
Participants were given a limited data set of recordings of players completing the challenge manually and up to four days of GPU time to train their AI model. Of over 660 participants in the challenge, not one was able to complete it fully. Some agents had learned how to navigate the world and even craft tools they'd need, but none succeeded in locating diamonds.
This challenge serves as an example of how Minecraft could be used to help teach young people about AI, and to let young people experiment with AI in an environment that appeals to them. It also serves as an example that simply throwing processing power behind an AI isn't necessarily the right way to develop one and shows the limitations that narrow data sets can have.
Source: BBC News
