Notes

An aI Learned to Play Minecraft and It's Pretty Good

Chess Go, Chess, and now-Minecraft. Artificial intelligence models have added a new victory to their gaming kill list.



- Off - English


With a mix of unlabeled Minecraft videos and a small set of those that were labeled by contractors, the artificial intelligence company OpenAI was able to train an artificial neural network to successfully play Minecraft, a significant important milestone for the technology, that had previously struggled to crack the game's simple and sloppy gameplay. Open AI engineers released their findings in the form of a blog post and paper this week.



The model of OpenAI could perform more than basic survival and crafting. It was able to perform many of the same complex tasks as human Minecraft player. OpenAI released a video of its model swimming, hunting and cooking animals. It even managed to master the "pillar jumping" technique. Deepmind was also able to train its MuZero AI to play Atari Games.



Previous AI models have famously relied on various types of reinforcement learning in the past to beat classic games such as Chess and Go. Minecraft however is a game that can be mastered by young children however, it presents challenges for AI systems because of its open-world structure and its open-ended nature.



There are plenty of videos on the internet about Minecraft gameplay. However, those videos only tell part of the story of how an AI learns how to play the game. According to OpenAI the flood of video data that isn't labeled excels at demonstrating "what" to do, but it doesn't provide exact key presses or mouse moments that are essential for an AI to know "how" to play.



This "how" problem was solved by engineers who developed a semi-supervised imitation learning process they call "Video PreTraining," also known as VPT. OpenAI basically gathered a brand smaller, but more detailed dataset of contractors that contained not just Minecraft gameplay but also examples of key press and other actions recorded by contractors. OpenAI created an additional model that utilizes videos of contractors to predict the next action in each step of the Minecraft film. Equipped with the basic knowhow the AI was able to successfully understand larger datasets of Minecraft videos online. Rather than just dump an inordinate amount of data onto their AI, the engineers took the time to teach it the fundamentals of inputs that are basic.



"For many tasks our models exhibit human level performance and we are the first to report computer agents that create diamond tools, which could be completed by skilled humans in upwards of 20 minutes (24,000 actions in the environment) of gameplay to achieve," OpenAI worte in their research paper describing the results.



All that training and contractor assistance reportedly resulted in a price that was around $160,000. According to ZDNet the bulk of that cash was given to contractors who made up around 4,500 hours of games. The hourly rate for contractors was $20.



You can see some footage of the AI cutting wood, directing its inventory, and searching caves for yourself below.
bagauc


If watching an AI that's worth the yearly salary of some surgeons play an 11-year-old indie game isn't exactly impressive It's worth taking a step back and seeing how far the technology has come. Three years ago the MineRL competition set teams of technologists the challenge of creating an AI capable of mining diamonds in Minecraft. This challenge was attempted by 660 contestants, with the majority of them failing. OpenAI's model is now capable of creating diamond tools.



OpenAI isn't just a technology company that relies on Minecraft for its AI experiments. In October, during its Build conference, Microsoft revealed a new AI Minecraft "agent" that operates within the game. When interfacing with Microsoft Minecraft agents, users can enter commands that are then generated automatically using the game's API. Wired informs us that users can type in phrases such as "come here" into the Minecraft bot and it will translate it into Minecraft code. The bot then walks forward. Apart from walking in the direction of Microsoft's Minecraft agent can also complete more complex tasks such as retrieving items from the game world and combining them to make something. It is able to do this quicker and more efficiently than the writer who has been a long time from his last Minecraft session.


Website: https://bagauc.com/