Dynamically learning efficient server/client network protocols for networked simulations .

摘要

With the rise of services like Steam and Xbox Live, multiplayer support has become essential to the success of many commercial video games. Explicit, server-client synchronisation models are bandwidth intensive and error prone to implement, while implicit, peer-to-peer synchronisation models are brittle, inflexible, and vulnerable to cheating. We present a generalised server-client network synchronisation model targeted at complex games, such as real time strategy games, that previously have only been feasible via peer-to-peer techniques. We use prediction, learning, and entropy coding techniques to learn a bandwidth-efficient incremental game state representation while guaranteeing both correctness of synchronised data and robustness in the face of unreliable network behavior. The resulting algorithms are efficient enough to synchronise the state of real time strategy games such as Blizzard's Starcraft (which can involve hundreds of in-game characters) using less than three kilobytes per second of bandwidth.