How amazing is the learning ability of artificial intelligence,
We can get a glimpse of the door through OpenAI, an AI-based experimental project Multi-Agent Hide and Seek.
This is the project primarily to investigate the ability of multiple agents to collaborate and compete in a virtual environment.
First, the tester created a virtual environment that included a labyrinthine scene with walls, obstacles, and hiding spots. In this environment, two groups of agents are placed: humans and ghosts.
Red is a person, blue is a ghost.
The goal of a person is to find and "catch" all the ghosts as quickly as possible, and the goal of the ghosts is to avoid being caught for as long as possible.
Using reinforcement learning algorithms, the agents begin training.
In the beginning, it was very easy for people to catch ghosts.
But after millions of rounds of learning,
Ghost constantly adjusts his strategy based on feedback,
Suddenly in a certain game, they learned to move the box to the entrance and lock it.
Once the box is locked, it cannot be moved, creating a closed environment for himself,
People can't get in.
So the ghost finally found the first strategy to win the game.
As the game progresses, after millions of rounds of evolution,
People finally found a breakthrough by using different algorithms and adjusting strategies,
They found that they could use the slope as a ladder.
So people learned to climb over the slope.
Jump into the wall to catch ghosts,
The AI playing the ghost can only continue to look for ways to win.
Soon they had a new plan,
Take away the ramp tools used by people,
So before closing the house,
They first moved the slope to their house.
And then close the gate.
Then people can't enter the house again.
To add to the challenge,
The tester put both parties into a more open environment.
At the same time, they are given more props to challenge.
At the beginning, ghosts were easy to find, and as the game continued to be repeated, they gradually learned to make a wall with wooden planks to surround themselves, and in order to prevent people from using the slope, they locked the slope in advance, seemingly impeccable plan, but did not expect that people actually learned new skills, they pushed the box to the edge of the slope, and then used the slope to put the box, and then controlled the box to drift to the edge of the wall, just like that, the person climbed over the wall again and caught the ghost.
This is an innovation that breaks through the human imagination, and no one could have imagined that artificial intelligence would use boxes to drift, which stunned all the testers.
In this multi-agent hide-and-seek project, agents constantly optimize their behavior by interacting and fighting with their environment. They improve their strategies through trial and error, and adaptation, resulting in more intelligent and collaborative behavior in a game of multi-agent hide-and-seek. This process demonstrates the ability of an agent to learn and evolve.
I dare not imagine that if they continue to learn like this, artificial intelligence will eventually evolve to what extent, and whether it will replace humans or even eliminate humans.