These New AI Bots Learn To Use Tools And Work Together To Win A Game

AI using tools is probably the next level of artificial intelligence. OpenAI researchers have shown the result of when they programmed AI that can learn from its mistakes while they played hide and seek.

The AI played the game in two teams with an environment featuring walls, ramps, and boxed which could be moved or locked. As soon as one bot has locked an item in place, the other team’s bot cannot move that item.

At its early stage, the bots learned to build forts using those boxes, walls, and ramps to hide from their opponents, and they often worked together for building the forts. Finally, the seeking team learned that the ramps could be used for getting over the walls of the forts. Then, the hiding team began locking the ramps before they hid. The seeking team then learned to get on top of the blocks using the locked ramps, they surfed to the forts then jumped in. As a result, then, the hiding team learned to lock the boxes too.

The team explained their development in a paper released earlier this month by OpenAI, which Elon Musk co-founded in December 2015.

Lead researcher Bowen Baker of the project said that the most surprising thing was how box surfing emerged.

The most entertaining part in this experiment is probably when the bots began finding the game’s glitches that they could use to win. For instance, the bots, at one point, figured that they could push the ramps through the outside walls at some angles to get rid of them.

All we’ve said took place in almost 500 million hide-and-seek games. The research team said that they consider it a truly interesting analogue to human evolution on the Earth.

According to Baker, the AI completed the games in five days, and at each time, there were 4,000 games being played simultaneously.

By using this method to teach artificial intelligence to use teamwork as well as learning from past mistakes to solve complex problems, the researchers can change how we think about the way to create advanced artificial intelligence. While this may only sound like children learning to play the game and learn to win it, this research can have huge implications.

It’s written in the paper that such an experiment can help illustrate how this way of developing AI can result in human-relevant and physically grounded behavior.

The way these bots learned was similar to how humans learn, which means this experiment could show how we can generate artificial intelligence mirroring human thinking. The AI was able to take advantage of the game’s flaws, meaning they could figure out solutions which humans might not have thought of. Baker said:

Back in April, according to Inverse, OpenAI made headlines as its bot team could defeat a pro sports team in Dota 2. It’s exciting to see its next creative project which can bring more potentials.

According to Baker:

However, one question remained is whether the company would ever put this development into questionable use, as it allayed in December 2015:

Nevertheless, Elon Musk said that he left the company at least one year before he said on Twitter that he no longer was on the board.