Modeling Approach

[NOTE: Freewriting first draft for the "Military Modeling" chapter of the upcoming book, Dynamic Modeling, edited by Paul Fishwick, to be published by CRC Press in 2006.]

In the previous section we discussed many patterns of relationships that exist between multiple models and talked in very general terms about what would be represented in those models. However, we did not explore specific mathematic or logical algorithms that would be used in those models. In practice, the number of techniques, algorithms, and equations that are used in military models is close to uncountable. It is not possible to describe all of them or even those that might be considered “the best”. So many different problems are studied with military models that there is no “best” approach that can applied universally when representing a specific vehicle, human, or unit. However, the techniques that are used do fall into distinct categories. In this section we will discuss four categories of modeling dynamics that are often used in military simulation systems.

Physics

Physics-based models are most often found in engineering and virtual simulation systems. For example, a missile pursuing a target would be represented by the physics of motion, momentum, mass, and aerodynamics. Changes in the fin positions would drive aerodynamic equations and change the vector of the missile based on the forces at work on the mass of the missile. Similarly, the seeker head in the missile would scan the environment electronically using the same pattern, revisit rates, and sampling rates of the real missile. This behavior would allow the simulated missile to collect data about a target in the same way that the real missile does.

Physics-based models are most often used to analyze the behavior of an existing weapon or to assist in the design of a new weapon. Understanding exactly how the pieces of the system will behave is an important part of exploring the design space to find optimum capabilities and combinations of capabilities that are optimum for the entire system.

Physics models require a great deal of data and mathematics. The data must be available for the system being modeled, the environment in which it is operating, and any other objects that it will interact with. Mathematics are required to represent a number of different behaviors of the system, interactions that occur within the system, and interactions that occur with other objects. Given this need, it is not sufficient to collect data and equations only for the missile that is to be studied. The model builders must do the same for the environment and for any objects that will interact with the missile.

Because of the volume of data, and the number and complexity of the equations that are required, physics models are necessarily reserved for smaller scenarios that involve only a few objects. Once constructed, the models can be computationally intensive. This means either purchasing a number of high-powered computers or accepting extremely long simulation times. The budget of the project limits the former and the schedule limits the latter. The project is literally a compromise of what the project can afford in time, money, and skilled staff. These limitations are one of the primary causes of the diversity in military modeling solutions. Constraints have forced generations of modelers to create unique representations of their problem.

Stochastic

Stochastic processes, probability and statistics, are most often found in virtual and constructive models. As simulation systems grow larger in their scope of representation, there is a need to capture many more activities and interactions in models. Lacking the detailed knowledge, breadth of expertise, access to data, time to build, and compute power to run a pure physics-based system, modelers have often resorted to a statistical representation of objects and interactions. In this case the models capture the behavior of many iterations of an event and represent individual event results using a probability function and the results of a pseudo random number generator. This type of modeling was introduced to the military modeling community by Stanislaw Ulam when he was working on the design of atomic weapons during World War II (REFERENCE). Ulam encountered a number of problems for which the specific physical behaviors were not known, but where the pattern of outcomes had been measured. Therefore, he chose to use the statistical properties of the event and rely on multiple simulation runs to arrive at an accurate behavior for the entire system.

The previous missile example lends itself well to stochastic models. Instead of representing all of the minute physical interactions, a modeler could choose to represent the outcome of a missile engagement given a limited number of input variables governing each event and recourse to a probability distribution. The use of a pseudo random number in decision-making means that no one engagement contains all of the details of the event as in the physics model above. However, if the model is run a number of times, the randomness of multiple models will blend together and arrive at an accumulated result that is representative of the system behavior that emerges from all of the interacting models.

Stochastic modeling has proven to be extremely useful because it allows modelers to study problems that were previously beyond the limitations of physics models. This has led to the creation of very large simulation systems capable of representing hundreds or thousands of events and objects on a battlefield. However, these models also require that their creators understand both the physical behavior of the system and the statistical aggregation of those behaviors in order to create accurate stochastic models.

Logical Process

Physics and stochastic models are not appropriate for representing the processing of information that is carried out in a computer. These activities are better represented as a sequence of logical steps that make up a defined process. Within the missile there are controllers and computers that process information and make specific decisions based on that stimuli. A model of the missile may best serve the needs of a study by replicating that logic to control the missile’s reaction to maneuvering targets or its response to control signals from an aircraft.

Logical models may also be used to capture the core rules of combat, or the steps that are followed by automated objects in carrying out their mission. These objects may be aircraft, ground vehicles, weapons, sensors, or any other battlefield object. When an object is controlled by a simulation system rather than a human operator, most of the time it is following a logical set of defined processes. These instructions tell it when to move, which direction to go, how fast to proceed, which objects to focus on, and which to ignore. These may be very complex processes, but they do not involve equations of physics or random decision points. In situations when an object should follow some form of “textbook” operation, the logical models are an excellent method of encoding this.

Finite State Machines (or Automata) (FSM) are often used to assist in organizing very complex sets of behaviors. FSM allow the modeler to capture hierarchical behaviors, triggers for changing from one behavior to another, encapsulated behaviors that can be reused in multiple FSM, and deterministic behavior that can be mapped and validated. Military systems that are known as Computer Generated Forces (CGF) or Semi-Automated Forces (SAF) systems often contain a large number of FSM logic models. CGF systems are used to provide automated control of several dozen or hundred objects. A human may provide the overall mission and direction, but the CGF will supplement this with detailed control of movement and engagement through the use of FSM. These systems are not limited to logical models, but may integrate models of all the types described in this section. CGF have proven extremely useful in reducing the number of humans necessary to control simulated battlefield activities by moving detailed control from the hands of the human controller to the FSM logic.

Artificial Intelligence

Artificial intelligence also encompasses logic process models like FSM and production systems, but it is broader than that. In military modeling, these techniques are used to represent the behavior of humans, groups, and objects that are controlled by humans. The focus is on replicating the decisions that are made under a specific set of stimuli. To accomplish this, modelers and researchers have turned to FSM, expert systems, case-based reasoning, neural networks, means-ends analysis, constraint satisfaction, learning systems, and any other technique that shows promise in accurately capturing the complex reasoning process of humans.

The missile guidance and navigation example that we have been using is not ideal in this area. Though a missile model may use a FSM to model its behavior, it is not attempting to create an artificial representation of intelligence, rather it represents a logical process that is followed robotically. If the missile were being controlled remotely by a human who is viewing the target on a computer screen, then the behavior of the human might be represented using an AI technique. A neural network may represent the human’s ability to discriminate a target in the scene and means-ends analysis may represent the humans decision process in selecting a target, leading its position, and switching from one target to another opportunistically.

AI techniques usually focus on processing information in a human-like manner. Using databases or rule sets, the algorithms attempt to make deductions that lead to behavior selection. The deductive process may be deterministic or stochastic (Russell and Norvig, 2000).