Using Context-Based Neural Networks to Maintain Coherence Among Entities' States in a Distributed Simulation

For entities to interact meaningfully in a distributed simulation environment, coherence among the entities' states must be maintained. Because continuous state updates for all entities in the simulation normally require large amounts of network bandwidth, motion equations (i.e., dead-reckoning models) are frequently used to reduce the number of communications updates. However, even with the use of such dead-reckoning models, networking and communications limitations still exist in currently fielded systems. An effective approach to reducing the communications requirements is achieved by replacing these predictive dead-reckoning models with neural networks. This paper presents the background and motivation for the research, the architecture and training algorithms of the networks, and the integration of the networks into a large-scale simulation environment. Quantitative measures from the experiments reveal that the use of neural networks can effectively reduce the number of communication updates required to maintain entity-state coherence. However, the neural networks may also be more difficult to scale than the currently used dead-reckoning algorithms.