Neural Network-Based Load Prediction for Highly Dynamic Distributed Online Games ⋆ Vlad Nae, Radu Prodan, and Thomas Fahringer Institute of Computer Science, University of Innsbruck Technikerstraße 21a, A-6020 Innsbruck, Austria {vlad,radu,tf}@dps.uibk.ac.at Abstract. We propose a neural network-based prediction method for the future entity layout in massively multiplayer online games. Our ser- vice has the potential to timely foresee critical hot-spots in fast-paced First Person Shooter action games that saturate the game servers which no longer respond to user actions at the required rate. Using our service, proactive load balancing (and redistribution) actions can be triggered. We show results based on a realistic simulation environment that demonstrate the advantages of our method compared to other conven- tional ones, especially due to its ability to adapt to different load patterns. 1 Introduction Online games can be seen as a collection of game servers that are concurrently accessed by a number of users that dynamically interact with each other within a game session. Clients connect directly to one game server, send their play actions (e.g. movements, collection of items) and receive appropriate responses. The responses must be delivered promptly within a given time interval to ensure a smooth, responsive and fair experience for all players. For the vast majority of games there is a similar computational model. The game server runs a large loop in which the state of all entities is first computed and then broadcasted to the clients. All entities within a specific avatar’s area of interest (usually a surrounding zone as shown in Figure 1(a)) are considered to be interacting with it and have an impact on its state. The more populated the entities’ areas of interest are and the more interactions between entities exist, the higher the load of the underlying game server is. An overloaded game server delivers state updates to its connected clients (i.e. movements and actions of teammates and opponents) at a significantly lower frequency than required, which makes the overall environment fragmented, unrealistic and unplayable. First Person Shooter(FPS) action games are considered as the next generation of distributed online games, for which no massively multiplayer support exists yet. Fast-paced online FPS games have very demanding Quality of Service (QoS) requirements in terms of client state computations and updates (i.e. around 35 ⋆ This work is funded by EU through IST-034601 edutain@grid and IST-004265 Core- Grid projects. E. Luque, T. Margalef, and D. Ben´ ıtez (Eds.): Euro-Par 2008, LNCS 5168, pp. 202–211, 2008. c  Springer-Verlag Berlin Heidelberg 2008