An Investigation of System Identification Techniques for Simulation Model Abstraction

Abstract

PubUc raporling burden lor this colectian of mlormation a estsnated to average 1 how par response, including tha tima for reviewing instructions, seerching oiisting data sourcts. gathering and maintaining Itta data needed, and competing and ravtawing lha coatction of information. Sand comments regarding this burdan estanata or any othtr aspect ol this collaction of intotmatwn, including suggasttons for raducmg this burden, to Washington Headquarters Services. This report summarizes research into the application of system identification techniques to simulation model abstraction. System identification produces simplified mathematical models that approximate the dynamic behaviors of the underlying stochastic simulations. Four state-space system identification techniques were examined: Canonical State-space, Compartmental Models, Maximum Entropy, and Hidden Markov Models (HMM). Two stochastic simulation models were identified: the "Attrition Simulation", a simulation of two opposing forces, each operating with multiple weapon system types; and the "Mission Simulation," a simulation of a squadron of aircraft performing battlefield air interdiction. The system identification techniques were evaluated and compared under a variety of scenarios on how well they replicate the distributions of the simulation states and decision outputs. Encouraging results were achieved by the HMM technique applied to Attrition Simulation -and by the Maximum Entropy technique applied to the Mission Simulation. This report also discusses the run-time performance of the algorithms, the development of suitable model structures, and implications for future efforts. SUBJECT TERMS AbstractThis report summarizes research into the application of system identification techniques to simulation model abstraction. System identification produces simplified mathematical models that approximate the dynamic behaviors of the underlying stochastic simulations. Four state-space system identification techniques were examined: Canonical State-Space, Compartmental Models, Maximum Entropy, and Hidden Markov Models (HMM). Two stochastic simulation models were identified: the "Attrition Simulation", a simulation of two opposing forces, each operating with multiple weapon system types; and the "Mission Simulation", a simulation of a squadron of aircraft performing battlefield air interdiction. The system identification techniques were evaluated and compared under a variety of scenarios on how well they replicate the distributions of the simulation states and decision outputs. Encouraging results were achieved by the HMM technique applied to the Attrition Simulation -and by the Maximum Entropy technique applied to the Mission Simulation. This report also discusses the run-time performance of the algorithms, the development of suitable model structures, and implications for future efforts.

A Simulation-optimization Approach to Air Warfare Planning

“…The attrition portions of the simulation (Phases 1 and 3) are designed as a multi-weapon type, stochastic Lanchester model. The state of each side (Red/Blue) is provided by a state-vector of weapon-system types (as in [31]) by location. Each unit on each side has a set of parameters designating the rates at which they can destroy opposing force types: λij = the rate (per planning period) at which forces of type i destroy opposing forces of type j when engaged solely against weapons of type j (assume that i and j incorporate the notion of force type, role, and location).…”