All for one one for All

Imagine a school where everyone tries to help the weakest student, and there is no competition among them except for being nice and supportive! Or companies who, instead of competing with their customers, try to make the world a better place! Yes, it is too big of a goal, but every big move starts from a small step. The small step in my thesis is working on the question "how to make groups work together in a better way" applied to the world of artificial intelligence. To train virtual agents to successfully work together, one must overcome the same problems humans face, such as selfishness. While agents might not be selfish, we typically optimize or preferentially select them on their individual performance. Thus, the selfish agents would also win again. To study and remedy this problem, I made a game for multiple artificial agents competing with each other. I also devised different objective functions so that selfishness might not be the optimal strategy. Specifically, Darwin's theory of Evolution and the group selection mechanism provides a different mechanism to incentivize groups to cooperate. The idea is that when evolving agents under these conditions, the last generation will include only cooperating agents. I used a framework that includes an evolvable neural network for implementing the agents and a genetic algorithm working based on specific rewarding schemes. I found that different rewarding schemes result in different agent behavior. Specifically, when agents evolve in groups whose performance is measured by their weakest member, they not only evolve to become an effective group, but they receive rewards more equally