| This dissertation develops and applies neural-network based movement models towards the problem of maintaining coherence among or “synchronizing” entities' states in a distributed simulation environment. This work has potential impact on a new form of distributed training in a live environment called embedded training, where network bandwidth for communications is a constrained resource.; The following two important issues are raised in this dissertation: (1) how to effectively model a near-term movement skill model from real-time data and how to measure the performance of these models, and (2) whether this approach will generalize to human driving. To these ends, a modeling scheme (i.e., features, representation schemes, network architecture, and sampling strategy, etc.) that results in the development of a neural-network based movement model is empirically developed and then the performance of these models are evaluated in a simulated environment. As part of this effort, the novel use of explicit, modular decomposition schemes are evaluated as a mechanism to improve model performance.; In consideration of the second issue, the fundamental principles of the previously determined approach are evaluated with actual human performance data. As part of this effort, the potential for generalization was evaluated by using Subject Matter Expert (SME) generated data for testing that was not used in the training of the neural-network based near-term movement models. Secondly, the potential for generalization was evaluated by applying the movement models developed from this SME SME-1 to a second individual, SME-2.; Finally, tangential to the primary objective of the dissertation, the use of the neural-network based movement model as a generator or controller of human skill was investigated. As such, its performance (i.e., fidelity with respect to human performance data and processing time requirements) was compared to the performance of the models used in prevalent semi-automated force (SAF) systems. |
