In a new article from NewScientist, a new AI program learned how to play and master an Atari computer game, Skiing, six thousand times faster than traditional methods. Older games are used as a benchmark to measure AI ability due to the simple controls of each game that in turn have greater complexity as levels are completed. According to the researchers, their approach allows AI to master Atari’s Skiing in just a few hours, compared to the weeks or even months it would take using conventional methods.
So, what’s the new method? Interestingly enough they taught the AI to read. The new technique allows AI to learn from the game’s original instruction manual, rather than through trial and error that have been commonplace for decades. In their experiments, the team used a machine learning algorithm called reinforcement learning to teach an AI agent how to play several classic video games, including Pac-Man, Space Invaders, and Montezuma’s Revenge.
Instead of relying on gameplay data, the AI agent was given access to the instruction manual for each game, allowing it to quickly learn the rules and objectives. The researchers found that this approach allowed the AI to learn the game much faster than traditional methods and achieve higher scores than human players in some cases.
The implications of this new technique and potential breakthroughs repeated are significant for the field of AI and machine learning. Not only does it provide a more efficient way to train AI agents for video games, but it could also have applications in other areas, such as robotics and autonomous vehicles which have traditionally relied on trial-and-error methods of learning.
One example of the application of this new technique has to do with autonomous cars. Imagine a self-driving car that can learn from a detailed instruction manual on navigating a specific city. Rather than relying solely on trial and error, the car could quickly learn the rules of the road and the best routes to take, allowing it to adapt to new environments much more quickly and effectively by inputting text data.
No matter the application, this new way of teaching AI could have serious consequences. It can revolutionize the way we teach machines to learn and interact with their environment, a major issue with machine learning models as they’re taken out of their control environments and are forced to interact with the real world and its variables. In short, what is taking years in teaching a model or AI program could easily become a fraction of the time.