To observe the divide, the best Artificial Intelligence (AI) and the mental capabilities of a seven-year-old child, there is one perfect avenue to observe all this: the popular video game Minecraft. A young human being can learn how to find a very rare diamond by watching a 10-minute video on YouTube.
Artificial intelligence (AI) is however not that close in that competition. But in a unique scenario at the end of the month, researchers are hoping to reduce the gap between the machine and the child in spotlight. For this purpose, they are reducing computing power need to train AI based system.
Competitors may take up to four days and use no more than eight million steps in training their AI systems in searching for the diamond. Still, that is longer than what it would take for a child to learn. However, this is much faster than typical AI models these days.
The contest was designed to help speed up advances in an approach known as imitation learning. It contrasts with a popular technique known as reinforcement learning. Here, programs try out a ton load random actions via trial and error method to hone in on the most suitable process.
How helpful has reinforcement learning been?
Reinforcement learning has helped generate recommendations for Netflix users, made ways to help train robotic arms present in factories, and has ousted human beings in video games and online gaming. It does however require a lot of time and computing power.
Attempts on using reinforcement learning for creation of algorithms in safely driving a car or win sophisticated games (like Go) have involved hundreds (and thousands) of computers in parallel function for running in a collective manner, simulations, worth centuries. This is something only corporations and governments with enormous treasuries can afford to do so (report Facebook hack).
Imitation learning can help improve the efficiency of the learning process. It does so by mimicking the way humans (and possibly AI algorithms) tackle tasks and solve them. The coding event known as MineRL Competition (pronounced as ‘mineral’), encourages contestants to use this technique in teaching AI in playing the game.
Reinforcement-learning techniques however would not stand a chance in this competition on their own, as stated by the head of the MineRL Competition’s organizing team. Working at any random event of time, an AI might succeed only in chopping down a tree (or two trees) in the eight-million step limit of the competition.
This is just one of the prerequisites needed for creating an iron pickaxe in mining diamonds in the game. Exploration is quite difficult and imitation learning does give the AI system a good precursor about the environment it is in.
The contest, which was sponsored by Carnegie Mellon and Microsoft among a host of many others hopes to create an impact beyond location of gems in Planet Minecraft. This can help inspire coders in pushing the limits of imitation learning in reducing the number of steps required for learning, something the event’s organizers hope to do so.
This kind of research can ultimately help train Artificial Intelligence in a way that it can interact with human beings smoothly in a wide array of situations and navigate environments easily, especially those strewn with uncertainty and complications.
A research scientist from Google DeepMind states that imitation learning is the very epicenter of learning and is key to development of intelligence. The same scientist is a member of the MineRL Competition advisory committee. Also, imitation learning helps humans and AI learn tasks easily without having to figure out the solution and without the need of making things from ground zero.
Why Minecraft of all platforms and games around?
The organizers of the competition believe that Minecraft is in particular as good as any existing virtual training ground. In fact, Minecraft’s players showcase intelligent behavior. In the game’s renowned survival mode, players must defend themselves against monsters, farm food and continuously gather materials in making structures and create tools.
New players on Minecraft must however learn the game’s physics, understand how how to make resources and tools out of existing materials. The game is popular because of the creativity it enriches in the players who have constructed blocky virtual versions of a wide array of objects, such as the Eiffel Tower, Taj Mahal, Brandenburg Gate and the like.