Enhancing Model Understanding through Static Analysis

Simulation is used increasingly throughout the sciences for many purposes. While in many cases the model output is of primary interest, often the insight gained by the modeler through the simulation process into the behavior of the simulated system is the primary benefit. This insight can come from the action of building the model as well as observing its behavior through animations, execution traces, or statistical analysis of simulation output. However, much that could be of interest to a modeler may not be easily discernible through these traditional approaches, particularly as models become more complex. Static code analysis techniques can reveal aspects of models not readily apparent to the builders or users of the models, even when applied to relatively modest simulation models. Using a commercially available static code analysis tool, we were able to find documentation errors in the published paper, “Redundancy in Model Specifications,” by Nance, Overstreet and Page [8]. This additional information about model properties is unlikely to be detected in executing the models and contributes to the insights gained by modeling a complex system. Introduction & Motivation It is often stated by users of simulation that the primary benefit is not necessarily the data produced by the model, but the insight that building the model provides. At the most recent Winter Simulation Conference (December 2005), Paul et al. discussed this [10], noting that “simulation is usually resorted to because the problem is not well understood,” and more often than not, the simulation is no longer of interest once the problem is fully understood. We believe the static analysis techniques such as those discussed here can be used to enhance their understanding and complement insights gained through model execution. Insights can arise from many different sources. One can be surprised to learn that one event causes another seemingly-unrelated event. One can also gain insight when something that is expected to happen does not occur. Sometimes events can happen with regularity or as clusters which may not be noticed by a modeler and may reveal important aspects of the simulated system. Often these facts are not immediately obvious, particularly in large simulations [9, 8]. Anecdotal reports from modelers support the frequent difficulty of detecting important aspects of their models which when pointed out are quite useful. Static analysis of models has the potential to help a model builder interactively explore different properties of the model such as causal relationship among model components. Automatically generated information about models has the potential of overwhelming users with too much information; using interactive static code analysis, it may be possible to offer the right information at the right time. For example, in an experience of Overstreet, a model-coder was studying the implementation of a model. Upon running it, it was noted that every event in one group occurred exactly the same number of times. Once this was observed, it became obvious that the model structure dictated this; however, having the right information at the right time could have enhanced understanding earlier than at the end of the implementation. As will be described shortly, in a project whose objective was to explore the ability of existing code analysis tools to provide model information to modelers, a static slicing tool was used to analyze a model [8] and errors were found in the published documentation of the model. Although the goal was not to check for errors, the tool clearly could be used to provide better documentation and hence a more complete understanding of the model. These analyses have other obvious uses including aid in debugging, verification and documentation. We, however, are primarily interested in how these analysis techniques can be used to help a modeler gain additional insights into models s/he is using or constructing.