Researchers Are Working On AI-Training For Playing Minecraft – Here Are The Reasons

Perhaps there’s more overlap between nuclear fusion and Minecraft than just the sheer amount of time you can spend on either. MIT Technology Review claimed that OpenAI, a non-profit AI research organization, has developed the most advanced Minecraft playing bot. It was carried out by a novel training process by evaluating over 70,000 hours of human playtime. Researchers predict bot’s achievements may one day bring in game-changing technology. It included self-driving automobiles and almost endless renewable energy resources. Despite the fact that it is now limited to manufacturing pixelated tools and buildings.

Researchers used a method called imitation learning to create the first bot capable of crafting “diamond tools,” an item in Minecraft that typically takes humans around 20 minutes and 24,000 actions to manufacture. Imitation learning, as its name suggests, needs an AI to study and eventually surpass hundreds of samples of human input. But reinforcement learning, another well-liked AI development strategy, takes a less methodical, more haphazard approach by relying on trial and error to learn.

Traditional problems with imitation learning stem from the fact that researchers typically need to manually label “each step,” as explained by Technology Review. For example, “performing this action causes this to happen, doing that action makes that happen, and so on”. OpenAI developed a distinct neural network called Video Pre-Training that is capable of handling the labeling process (VPT). Researchers hired gig workers to play Minecraft for 2000 hours and record every keypress, and video to teach an AI bot.

By incorporating VPT, the new AI software was able to accomplish feats in Minecraft. Bots were solely dependent on reinforcement learning and had not been able to. For example, the approximate 970-step process of constructing a table from manufactured planks. The bot was able to handle building jobs with over 20,000 consecutive operations when emulation and supervised learning were coupled.

Imitation learning gains seen in Minecraft may one day benefit nuclear fusion research and self-driving automobiles. However, these are still far off. Emulation and reinforcement learning’s efficacy and ethics are still unknown.

In 2015, Elon Musk and Sam Altman were among the company’s co-founders. Peter Thiel was an early investor. However, Musk left the governing board this year in 2018.

Read Also: 11 Best Minecraft Prison Servers