Martin Zinkevich
Abstract: Imagine a game where the constant strategy beats out the most popular multi-agent strategies from the past fifty years. How do we learn from this? Like Axelrod's original iterated Prisoner's Dilemma competition, the Lemonade Stand Game Competition led us to question the fundamental aspects of learning in games. Specifically, convergence to equilibria took place far faster than simple regret minimizers would achieve. This competition has shown that, with knowledge of the utilities, bots can be constructed that do not just learn over a lifetime how to cooperate, but cooperate in a few moves. However, it is still beyond our capacity for humans to algorithmically formalize how they set such objectives for an arbitrary game.
In this talk I will discuss some of the results from the four competitions we have had, and present directions for the competition to follow in the following years.