OpenAI Trains Minecraft Bot to Play 70,000 Hours

The most recent artificial intelligence study from OpenAI demonstrates that a skilled Minecraft computer can outperform people.

OpenAI focuses on machine learning and artificial intelligence (AI) developments that benefit humanity. A machine that can play Minecraft has been successfully taught by the company using more than 70,000 footage of actual gaming.

The achievement is more than simply a game-playing robot; it represents a significant advance in machine learning techniques based on observation and imitation.

OpenAI Trains Bot To Play Minecraft for 70,000 Hours

The researchers want to achieve for imitation learning what GPT-3 accomplished for big language models by utilizing this resource. According to Bowen Baker of OpenAI, a team member responsible for the new Minecraft bot, "in the last few years we've seen the rise of this GPT-3 paradigm where we see amazing capabilities come from big models trained on enormous swathes of the internet."

The issue with current imitation learning methods is that each step of the video demonstrations needs to be labeled: taking this action causes this to happen, doing that action causes that to happen, and so on. Since this type of manual annotation requires a lot of labor, such data sets are frequently tiny. The millions of films that are available online were the source of the new data collection that Baker and his colleagues were trying to create.

ALSO READ: How Metaverse Rules The Gaming World?

How Experts Trained the Bot To Play Minecraft

The team's method, known as Video Pre-Training (per Parametric Architecture), circumvents the imitation learning bottleneck by teaching a different neural network to identify films automatically. The initial step was for the researchers to recruit crowdworkers to play Minecraft while they logged keyboard and mouse actions along with screen captures. As a result, they had access to 2,000 hours of annotated Minecraft gameplay, which they utilized to train a model to correlate user actions with visual results. In certain circumstances, pressing a mouse button causes the character, for instance, to swing its ax.

The next stage was to train the Minecraft bot on this bigger data set by using this model to create action labels for 70,000 hours of unlabeled footage that was downloaded from the internet.

"This work is another testament to the power of scaling up models and training on massive data sets to get good performance," says Natasha Jaques, who works on multi-agent reinforcement learning at Google and the University of California, Berkeley (per MITNews).

It is unlikely to happen any time soon, according to Matthew Guzdial from the University of Alberta in Canada, who has taught AI how to play games like Super Mario Bros. using films.

Actions are carried out in video games like Super Mario Bros. and Minecraft by pressing buttons. Actions in the physical world are significantly more complex and difficult for robots to learn. Guzdial claims that it offers up a ton of brand-new research questions.

Check out more news and information on Technology in Science Times.