[ad_1]
By
OpenAI
Never play with a bot, he will find a way to cheat if he can. An OpenAI team, an artificial intelligence lab co-founded by San Francisco-based Elon Musk, has developed intelligent artificial robots that have learned to cooperate by playing hide-and-seek. Robots have also learned how to use basic tools and that defying the laws of physics can help you win.
In April, a team of robots called OpenAI Five defeated the world champions in the team video game. DOTA 2. Hidden robots use similar principles to learn, but the simpler game allows a more inventive game.
Bowen Baker of OpenAI and his colleagues wanted to see if the team dynamics of the OpenAI Five could be used to generate skills that might one day be useful to humans. "It's hard to imagine a DOTA arise from a game and solve real problems, "he says.
Publicity
The researchers released their robots in a simulated environment filled with fixed walls and moving boxes and let them play hide-and-seek team games millions of times. The robots each had their own vision of the world and could not communicate directly with each other.
At first, the hiders simply ran away. But they soon realized that the quickest way to stop researchers was to find objects in the environment to hide, using them as a kind of tool. For example, they learned that boxes could be used to block doors and build simple hiding places.
The researchers learned that they could move a ramp and use it to climb walls. The bots then discovered that being a team player – passing on items or collaborating on a hiding place – was the fastest way to win.
The henchmen also learned to sabotage the other team, for example hiding the ramp before hiding. "When a team learns a new strategy, it creates extra pressure for the other team to adapt," Baker says. "It has this really interesting analog to how humans have evolved on Earth, where you've had constant competition between organisms."
But the real surprise came when robots began to exploit problems in the physical simulation of their environment. The researchers found that if they pushed a ramp towards a wall, they could launch themselves into the air and spot intruders from above. The adventurers discovered that they could permanently get rid of the ramps by passing them through the outer walls at a certain angle.
Such hints show that AIs are able to find solutions that humans lack, says Baker. "Maybe they will even be able to solve problems that humans do not yet know how to do."
However, this is a huge leap from virtual concealment to real problem solving. "The main limitation of this type of work is that it is in simulation," says Chelsea Finn of Stanford University in California. Earlier this year, Finn and his colleagues built a robotic arm that taught them to use objects placed in front, as tools, using a sponge to eliminate crowding, for example.
More on these topics:
[ad_2]
Source link