Proxels Applied to Sensitivity Analysis and Optimization of Discrete Stochastic Models

Simulation-based optimization or parameter tuning of discrete stochastic models becomes necessary, when no analytic expression for the goal function is available. Sensitivity analysis on the other hand is used to determine the required degree of detail that is needed when building a model. Both of these tasks are similar in the sense that one needs to repeatedly simulate a model with only slight changes in the tested parameter sets. Usually the steady state results using these slightly different parameter sets are also close to each other in some sense. This paper shows that state space-based simulation can reuse old simulation results to arrive at a steady state solution faster than when starting from an initial model state, if the successive model configurations are in some sense close to each other. Performance gains can be generated for sensitivity analysis and optimization compared to restarting the simulation every time anew. The paper will present the idea and an algorithm for performing successive runs of the Proxel simulation using the information from the previous ones. Experiments test the applicability to sensitivity analysis and optimization, and investigate the possible saving dependent on the size of the perturbation.