DeepMind, Alphabet Inc’s artificial intelligence unit, developed a research in which they investigated the conditions in which both humans and robots would choose to cooperate rather than compete. This research might be used to find out how we can prompt computer intelligence to manage complex systems, such as an environmental policy, traffic flows, or economy problems.
Joel Leibo is the lead author of the study in question. He revealed that the choice of cooperation or competition is dependent on the environment in which the agents operate. So far, this research cannot be applied in the real world, but it can help the researchers at DeepMind develop artificial intelligence that can function in imperfect environments.
Thus, in the future, the newly developed agents would be able to manage a world full of possibilities with both human and machine entities. They might be able to tackle different fields, from transport networks to stock markets.
The researchers used two games to see how machines learn to compete or to cooperate. The first game implied two software agents that had to maximize the number of apples that could be gathered in a two-dimensional digital space.
When the apples were scarce, the agents learned more quickly to attack one another by sending rays which could immobilize their opponents. When the apples were plentiful, they could learn to cooperate and co-exist peacefully.
However, when they tried more advanced machines that mimicked certain parts of the human brain, they discovered that they behaved more competitively regardless of how scarce of abundant the apples were. This might not be the answer we wanted to hear.
Wolfpack, the second game, made the agents play wolves that learned how to catch their prey. When their catches were successful, not only the predator, but all wolves present in the vicinity of the capture were rewarded. This made the AI agents learn to better cooperate. Unlike in the apple game, the more advanced the machines were, the more quickly they learned to cooperate.
The researchers explained the situation as following. In the apple game, the more complex activity involved was the attack. In Wolfpack, the most complex activity was cooperation. Therefore, the machines learn quickly what is more complex.
At the moment, we have no reason to worry. The less complex machines encountered difficulties in learning the complex processes. Therefore, until incredibly sophisticated artificial intelligence is developed, we should not worry that robots will try to conquer our world.
Image Source: Pixabay