Multitrajectory simulation performance for varying scenario sizes

Abstract

Multitrajectory Simulation allows random events in a simulation to generate multiple trajectories, a technique called splitting, with explicit management of the set of trajectories. The goal is to gain a better understanding of the possible outcome set of the simulation and scenario. This has been applied to a prototype combat simulation, eaglet which was designed to have similar, but simpler, representations of the features of the Eagle simulation used for Army analyses. The study compared the number of multitrajectory simulation trajectories with numbers of stochastic replications to experimentally determining the rate of convergence to a definitive outcome set. The definitive set was determined using very large numbers of replications to develop a plot of loss exchange ratio versus losses of one side. This was repeated with scenarios of from 40 to 320 units. While the multitrajectory technique gave superior results in general as expected, there were some anomalies, particularly in the smallest scenario, that illustrate limitations of the technique and the assessment method used. 1 BACKGROUND The goal of multitrajectory simulation is to explore the outcome space of a simulation, that is, the set of all possible outcomes, more systematically and less expensively (for a given quality of understanding) than can be achieved with conventional stochastic simulation. This may be considered a variance reduction technique, but the analysis goals may be formulated not only in terms of better estimates of statistical properties of the outcome set, e.g. a mean and variance for Measures of Effectiveness (MOEs), but also representative instances of extreme behavior or other interesting cases (Al-Hassan, Gilmer, and Sullivan 1997). The heart of the proposed method is to explicitly track each possible trajectory, as illustrated in Figure 1. When an event that would normally be stochastic occurs, instead of one outcome, multiple outcomes are generated, each constituting a trajectory having its own state. Because the trajectory bifurcates, this is also referred to as splitting, with cloning of the state. In concept, such a multiple trajectory simulation is integrated with its support system in such a way that its use provides outcomes with probabilities associated with each, an accounting for the key events or circumstances leading to the differences, and some measure of confidence in these results.