I am talking about relationships between AIs (e.g. 2 of them forming a couple, 3+ in family like relationship).
What knowledge could come out of such experimentation?
One example might be self-play in games. Since neural networks and deep learning depend on massive amounts of data, one way to generate data is to have two virtual machines play each other and record the experience. An example discussion can be found at http://www.cs.cornell.edu/boom/2001sp/Tsinteris/gammon.htm which uses reinforcement learning. I believe AlphaGo also uses this technique of self-play, and uses two independent neural networks, one reducing the search space and the other deciding on the best move in the remaining space, that in a sense cooperate to decide on the next move.