Notes

An AI Learned To Play Minecraft, And It's Actually Pretty Good

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



OpenAI, an artificial Intelligence company, was able train a neural system to play Minecraft. The technology had previously struggled to understand the game's simple and loose gameplay. Open AI engineers shared their findings in a research paper, and subsequently, a blog post.



OpenAI's model was capable of performing many complex tasks that a human Minecraft player would, beyond basic crafting and survival. OpenAI shares a video of the model swimming, hunting, cooking and eating animals in a blog post. It even mastered the game's "pillar jumping". Deepmind was also successful in training its MuZero AI how to play Atari games.



In order to beat classic games such as Chess and Go, previous AI models relied on reinforcement learning. Minecraft, on the other side, is intuitive enough to be mastered by young children, but presents a challenge for AI systems due its open-world and open-ended structure.



Although there are seemingly endless videos of Minecraft gameplay floating around the internet, they only tell part the story of how an AI learns to play the game. OpenAI claims that the unlabeled video data does a good job of demonstrating "what" but doesn't give precise key presses or mouse moments necessary for an AI understanding "how" to play.



Engineers solved the "how" problem by creating a semisupervised imitation learning method called "Video PreTraining," otherwise known as VPT. OpenAI essentially gathered a new, smaller dataset from contractors which included not just Minecraft gameplay but also examples of key presses and other actions recorded by the contractors. OpenAI then created another model which uses the contractors' videos to predict what action will come next in each step of a Minecraft video. Their AI was able to understand large amounts of Minecraft videos online once it had the basics. Instead of just dumping a lot of data on their AIs, engineers took the time teach them the basics of inputs.
Servers


"For many tasks our models exhibit human level performance, and we are the first to report computer agents that can craft diamond tools, which can take proficient humans upwards of 20 minutes (24,000 environment actions) of gameplay to accomplish," OpenAI worte in their research paper detailing the findings.



According to ZDNet, the cost of all that training and contractor assistance was approximately $160,000. Most of that cash, according to ZDNet, went to paying out the contractors who collectively assembled around 4,500 hours of gameplay. The contractors were paid $20 an hr.



Below you will see some footage showing the AI cutting wood, managing inventory, and searching caves.



If it seems a little strange to watch an AI that is essentially worth the annual salary paid to surgeons play an 11 year-old indie video game isn't impressive enough, it's worthwhile taking a step back to see how far tech has come. Three years ago, teams made up of technologists were given a simple task: to create an AI capable of successfully mining a Minecraft diamond. This challenge was attempted by 660 contestants. Every one of them failed. OpenAI's model can now create diamond tools.



OpenAI isn’t the only tech company using Minecraft to test its AI ideas. Microsoft unveiled an AI Minecraft agent last month at its Build conference. Users interacting with Microsoft Minecraft agents can type in commands that are then auto-generated using the game's software API. Wired notes that in practice, this means that users can enter a phrase like "come Here" and the Minecraft bot will translate it into Minecraft code. Microsoft's Minecraft agent is capable of performing more complicated tasks than simply walking. For example, it can retrieve items from the game and combine them to make something. And look, it can probably do that better and faster than this writer, who's several years removed from his last Minecraft session.


Read More: https://www.addictgaming.com/