Citation Context

...ance measure of effectiveness is the “price of anarchy” [1], which is the ratio of the productivity of an idealized centralized system to that of the decentralized one. However, the system outcome is binary in many situations: the team either succeeds or fails. The resulting price of anarchy could then be only 1 or ∞, which is too crude a range. Instead, let α (and β ) be the resource level (here: number of agents) needed for success with centralized (and decentralized, respectively) control. Then, we define β/α to be the price of decentralization. We use the popular game StarCraft as testbed [2] because properties like robustness and adaptability to change can only be evaluated experimentally, not analytically. StarCraft is a well-known preexisting domain with an API and all the complexity we need. Creating our own experimental testbed would inevitably be biased. Decentralized StarCraft agents need the capability to coordinate autonomously with each other, for example in the context of the self-assembly of teams of agents of minimum acceptable team ∗The research at Georgia Tech was supported by NSF Grant CMMI 1335301 and a David M. McKenney Family Professorship. The research at USC w...