Improving Scalability and Dependability of Stochastic MAS Analyses

When studying stochastic multiagent system models, it is currently most common to perform analysis using Monte Carlo simulations. However, this approach can be prohibitively expensive in certain settings. When system behavior is highly variable, a large number of simulations is needed to understand its behavior for even a single parameter configuration. Simulation performance can also scale poorly with certain parameter values, such as the number of agents and time steps. Working over a parameter space increases costs even further. We present an analytical approach to characterizing stochastic multiagent systems with (a) runtime independent of system variability, logarithmic in the number of time steps, and dependent on interaction size rather than population size; and (b) hard, non-probabilistic, error bounds. This method is applied in a sociodynamics setting, illustrating how an analytical approach can produce exact predictions quickly when a simulation approach could perform poorly. We also demonstrate how to characterize behavior across a parameter space and perform approximate inference and optimization tasks when using this analytical approach.