estadisticas web Skip to content

Collaborative or antagonistic? DeepMind tests the instinct of AI agents

Google DeepMind AI

if different types of artificial intelligence collide with each other, could it become a problem? Google DeepMind has answers with games.

A study conducted by the DeepMind division of Google shows that Artificial Intelligence (AI) would be able to do almost everything: from dreaming to creating artificial languages. This technology just keeps getting smarter, almost more than us. The fabric of our digital society is increasingly managed and controlled by the AI ​​whether it's controlling the traffic lights of a city, a company, a running financial system or the army American. But if different types of artificial intelligence collide with each other, could it become a problem?

The most scrupulous ask themselves: will these AI agents be mutually correct? What could happen in the event of system conflicts?

The DeepMind team searched for answers, through a study that they put together a series of social dilemmas, asking AI agents to find solutions. In practice, situations have been created that are able to test the way in which the agents could have posed in the event that they found themselves able to betray each other against a reward.

In a post in DeepMind's blog, it was shown how researchers tested AI in hypothetical situations by means of two video games.

In the first game he calls Gathering, two agents had to collect apples from a pile, both of which had the possibility of targeting the opponent with a laser beam, temporarily removing it from the game so as to ensure that they could collect more apples.

In second game Wolfpack two agents have to hunt a third in an environment full of obstacles, the points are obtained not only when an agent captures the prey, but also when all the other agents are close to the prey and when it is captured. What the researchers found rather interesting, but perhaps not surprising: the agents changed behavior becoming both cooperative and antagonistic, based on the objective.

For example, in the Gathering game when the apples were gathered in abundance, the agents were not interested in hitting the opponent with the laser, but when supplies were scarce the number of rays to hit increased, increasing the competitiveness between the two agents. Even more interesting than when an agent was more powerful in the calculation, he increased the frequency of the rays, while the weaker agent – regardless of the number of apples – was less aggressive.

The de DeepMind researchers hypothesize that the increase of rays in behavior by the most powerful, and therefore by advanced AI, was simply due to the fact that making an opponent's tracking was more demanding and computationally intense. The agent had to point his weapon at the other player and monitor the movements, activities that require more computing power. In this way, more powerful computing power was given to the more powerful artificial intelligence, allowing more solutions to be devised even at the cost of wasting energies of which, however, at the same time, insisting with an adversary by itself made cooperation more advantageous, even only in terms of resource savings.

On the contrary, in Wolfpack, the most powerful and advanced AI agent cooperated with the other agents. As the researchers explain, this is because learning to work with the other player to track the prey requires more computing power.

Google DeepMind AI

The results of the DeepMind study have shown how the behavior of AI agents is modified according to the rules who is facing. If these rules reward aggressive behavior, such as in Gathering, the agents will be more aggressive, but if the rules of the reward involve cooperation, then the system will be more cooperative.

This means that part of the challenge in controlling AI agents in the future will be to make sure that the right rules have been set up properly. But if artificial intelligence becomes more and more complex, and more and more integrated with our complicated digital fabric, it could be an impossible challenge to control our creations for a good purpose. But then what would happen?

The researchers wanted to emphasize that as a result of this research we might be able to understand and control complex multi-agent systems such as those applied in economics, traffic systems, or ecology to get a better planet. All these systems depend on our continued cooperation ".