Minecraft may not seem like a tool that can support advanced AI research. However, recent research from OpenAI shows that a well-trained Minecraft bot with artificial intelligence can be more successful than humans.OpenAI is a company focused on artificial intelligence (AI) and machine learning advances that benefit humanity. Recently, the firm announced that it has successfully trained a bot to play Minecraft using over 70,000 hours of gameplay videos. Success marks more than just a bot playing a game, it marks a big step forward in observation and imitation-based machine learning.
Significant advance for artificial intelligence education
The bot developed by OpenAI is a near-perfect example of imitated learning or supervised learning. Unlike learning reinforcements, which promise rewards when the goal is achieved, imitated learning is accomplished by watching people perform specific tasks. These two learning tactics apply to us humans as well. In this case, OpenAI leveraged existing gameplay videos and tutorials to teach its bots complex in-game moves that would require around 24,000 individual actions for a typical player to accomplish.
The biggest shortcoming of imitated learning is whether there are sufficient data sets and whether these data sets are of sufficient quality. On the other hand, the context of the inputs during the training and what the outcome is should be labeled. Therefore, this approach requires a little more effort, but this area is very important for systems that we want to specialize in a specific job.OpenAI, on the other hand, used a proprietary approach known as Video Pre-Training (VPT) to increase the number of tagged videos available, instead of dealing with the manual data tagging processes mentioned above. Initially, the system was trained with 2000 hours of explanatory Minecraft gameplay. The data from this process was then used to tag 70,000 hours of previously untagged Minecraft content. So this process gave the Minecraft bot a much larger dataset to analyze and emulate.
19530
