Short Bytes: A team of researchers at DeepMind has shown the behavior of AI agents with the help of two games Gathering and Wolfpack. The agents showed aggressive and cooperative behavior traits depending on the situation of the game. Also, the need of the hour is to pay attention while building such system than repenting later.We want to make artificial intelligence comparable to what humans have achieved in the thousands of years of their existence on the earth. Every other day, there is a new AI story making news around the globe. But the fact, AI might become inhuman in future, has also been acknowledged by the researchers and scientific minds, including Stephen Hawkings.
A research conducted by DeepMind throws light on the attitude developed by an artificial intelligence agent when it comes to working in shared environments with other agents. They set up a game called Gathering for two AI systems. The goal was to collect the maximum number of virtual apples. Also, the agents were equipped with laser beams to make each other dysfunctional for a few moments and increase their apple count in the meantime.
All went in harmony until the apples were in good quantity. As soon as their number started depleting, the agents started showing aggressive attitude in a quest to increase their count – they started laser-tagging each other.
In the case of larger and more complex agents, the frequency of laser tagging was higher than the simpler agents. The team of researchers notes that such agents “behave less cooperatively – no matter how we vary the scarcity of apples.”
The researchers analyzed agent behavior with another game called Wolfpack. In this, two agents posed as wolfs had to work in coordination to attack a third agent acting as the prey. The results, in this case, were in contrast to the Gathering game.
In Wolfpack, the agents had to work together to earn better rewards. Individual efforts could give them the prey. But it would be hard to protect the dead prey from scavengers, unlike, in the case, where the two catch prey together. You can read more in the research paper by the team.
The main point to be considered here is that the agents have to learn that their attitude and behavior should be according to the situation presented to them. In the Gathering game, a more self-centered and aggressive attitude helped them taste success while in Wolfpack, cooperation was the key. For the development of such artificial agents, special attention should be paid during the developing phase, rather than working on protective measures in the future. It is because AI is going to be an important part of our lives in the coming years.
Did you find the story helpful? Drop your thoughts and feedback.
Bonus: Simone Giertz “Shitty Robots” Compilation