The World’s Best AI Can’t Seem To Figure Out Minecraft
Microsoft oversaw a challenge that asked for coders to submit programs that can mine diamonds in Minecraft, but none succeeded.
You Are Reading :The Worlds Best AI Cant Seem To Figure Out Minecraft
From self-driving cars to chess-playing computers, humans are already taking a backseat to AI capability. There is one area, however, where artificial intelligence hasn’t yet caught up: Minecraft.
Last summer, Microsoft oversaw a challenge that asked for coders to submit programs that can mine diamonds all on their own. Close to 700 entries were submitted, but not a single one could perform the assigned task.
The results of the competition, which was dubbed MineRL (pronounced mineral), were officially announced this weekend at the NeurIPS AI conference in Vancouver. Even though there were no “winners,” the competition organizers were still pleased with the results.
Katja Hoffman, the principal Investigator at Microsoft Research, said in a statement, “The task we posed is very hard. Finding a diamond in Minecraft takes many steps—from cutting trees, to making tools, to exploring caves and actually finding a diamond. While no submitted agent has fully solved the task, they have made a lot of progress and learned to make many of the tools needed along the way.”
Minecraft is one of the most wide-reaching and accessible games in video-game history. For most players, whether you’re ten or thirty years old, mining diamonds is a task that takes no more than half an hour to master. If you want to speed up the process, there are plenty of step-by-step guides and YouTube tutorials to show you the way. But, for a computer, even one of Minecraft’s most basic activities is incomprehensible.
Machine learning takes place in one of two ways: imitation learning and reinforcement learning. Imitation learning asks for programs to watch a task and then recreate it through, well, imitation. Reinforcement learning takes place in a built environment, within which a program will teach itself how to complete a task through trial and error. For difficult tasks, AI usually combines these two methods of learning in order to fully learn a new skill. Some researchers suggest that reinforcement learning on its own is superior for teaching AI new skills, but it takes a massive amount of computing power and is only accessible to well-funded labs and bigger organizations.
Related: The Vatican Has Its Very Own Minecraft Server
As reported by The Verge, limited resources were available to MineRL participants. Some of the most advanced programs in the world took years to fine-tune. In comparison, MineRL entries had just four days to learn how to mine diamonds. As a result, many of the programs were able to complete tasks beyond the individual stages that are required in the process of diamond mining. For instance, some were able to collect wood for a pickaxe but couldn’t string multiple steps in the process together.
Lead researcher and organizer of the competition, William Gus, says that the time constraint and lack of resources that are afforded to competition participants are meant to demonstrate a point.
“Throwing massive compute at problems isn’t necessarily the right way for us to push the state of the art as a field,” he tells the BBC. “It works directly against democratizing access to these reinforcement learning systems, and leaves the ability to train agents in complex environments to corporations with swathes of compute.”
However, the competition’s lack of promising solutions may suggest that larger agencies like Google’s DeepMind may be the best place to turn.
Source: The Verge
Link Source : https://www.thegamer.com/the-worlds-best-ai-cant-seem-to-figure-out-minecraft/