As artificial intelligence become stronger and more sophisticated, there is a need to further test its behaviors in a given situation. The scientists of Google's DeepMind tested it and found out that AI has the tendency to become aggressive, especially when pitted with another AI.
Since artificial intelligence is self-learning, DeepMind scientists want to know if they behave the same way humans do in certain social situations. In order to find out, they subjected two AIs to compete with each other in two games called Gathering and Wolfpack assuming the role of an agent.
In Gathering, the objective of the game was for both players to collect apples found in the middle of the screen. They are rewarded a point is added for each apple they collect. They can also shoot each other where the loser disappears from the screen temporarily.
The game went quite peacefully at the beginning of the game with both AIs only picking apples. However, when the number of apples dwindled, that's where the competition began as the AIs began to shoot at each other. The researchers noted that greed is a powerful motivation for the AIs to eliminate their opponent and collect all the apples.
In Wolfpack, the AIs take the role of wolves out to hunt and kill their prey. The AIs are taught that the prey is very dangerous that although one wolf can capture it, there is also a risk of losing the prey to other scavengers. On the other hand, they have a higher success rate when they work to protect the carcass together against other scavengers.
The researchers noted that it is also possible for both AIs to work together harmoniously if the outcome benefits both parties in a similar level.
With this discovery, Google scientists hope to develop a more complex multi-agent system that would learn to cooperate with each other.
