Paul K. Davis, RAND and the RAND Graduate School
Donald Blumenthal, Gualala, California
Donald Gaver, Naval Postgraduate School
The design of JWARS and other new combat models should raise numerous issues about modeling approach and phenomenology. To some extent this has happened, particularly with DOD's recognition that such next-generation models must represent the effects of the C4ISR systems on which much modern defense planning is focusing. In many other respects, however, discussions to date have not converged and have too often been conducted at the level of “labels” used as litmus tests. Some of the labels dividing people in discussion include Lanchester models, attrition models, deterministic models, and configural theory. There have been numerous heated discussions on such matters because of the “Grand Canyon” that separates the domains of modelers and analysts working at different levels of resolution and, typically, on different types of problems. In this appendix we try to shed some light on the issues. Readers should understand, however, that there are chronic controversies on these matters, and no two authors are likely to emphasize the same issues. Although our examples pertain mostly to ground combat, the principles involved apply also to naval and air warfare.
Despite the hundreds of papers written about them, Lanchester equations (as most people understand this term) are largely irrelevant to today 's combat modeling by DOD, which uses computer simulations, not simplistic constant-coeffi-
Intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text on the opening pages of each chapter. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.
Do not use for reproduction, copying, pasting, or reading; exclusively for search engines.
OCR for page 226
Technology for the United States Navy and Marine Corps, 2000-2035: Becoming a 21st-Century Force I Combat Modeling Issues Paul K. Davis, RAND and the RAND Graduate School Donald Blumenthal, Gualala, California Donald Gaver, Naval Postgraduate School INTRODUCTION The design of JWARS and other new combat models should raise numerous issues about modeling approach and phenomenology. To some extent this has happened, particularly with DOD's recognition that such next-generation models must represent the effects of the C4ISR systems on which much modern defense planning is focusing. In many other respects, however, discussions to date have not converged and have too often been conducted at the level of “labels” used as litmus tests. Some of the labels dividing people in discussion include Lanchester models, attrition models, deterministic models, and configural theory. There have been numerous heated discussions on such matters because of the “Grand Canyon” that separates the domains of modelers and analysts working at different levels of resolution and, typically, on different types of problems. In this appendix we try to shed some light on the issues. Readers should understand, however, that there are chronic controversies on these matters, and no two authors are likely to emphasize the same issues. Although our examples pertain mostly to ground combat, the principles involved apply also to naval and air warfare. MAJOR OBSERVATIONS Lanchester Equations as Red Herrings Despite the hundreds of papers written about them, Lanchester equations (as most people understand this term) are largely irrelevant to today 's combat modeling by DOD, which uses computer simulations, not simplistic constant-coeffi-
OCR for page 226
Technology for the United States Navy and Marine Corps, 2000-2035: Becoming a 21st-Century Force cient differential equations such as the Lanchester-square-law. 1 Lanchester equations will probably remain quite useful for making particular points in the classroom (e.g., illustrating the power of concentration or the value of “crossing the T” in classic naval engagements) or theoretical papers, but to argue about their more general validity is to chase red herrings. It is the simulations, not the Lanchester differential equations, that should be examined. Today's higher-level combat simulations (e.g., those at division, corps, and theater levels) are best seen as implementing aggregate state-space models (something much broader than Lanchester models). The basic notion is that the “state” of the system (the two opposed forces, their strategies, and the environment in which they fight) can be represented by a collection of variables such as counts of personnel and vehicles in an area, and terrain factors characterizing that area, rather than the locations and current behaviors of all the individual entities such as individual soldiers and tanks. Usually, the simulation then generates the predicted future state as a function of the current (aggregate) state. In more general formulations, there can be “memory effects” of previous states as well. Again, the variables affecting this prediction are not just the sides' strengths (much less their scalar strengths, as in the simpler Lanchester equations). Instead, the predicted change of state depends on many other factors such as terrain, defender preparations, flank exposure, strategy, and tactics. One important change of state, typically made at the end of time periods or when some significant event occurs, is a change of strategy or tactics (e.g., a decision to attack or withdraw, or to maneuver reinforcements to a trouble area). It is then true that the close-combat ground-force attrition in a given time step is sometimes approximated by a local use of some Lanchester equation, but the “coefficients” used can be highly situation dependent, that is, dependent on many other state variables that change over time (Allen, 1992, 1995). Thus, the simulation does not (or at least is not intended to) behave like a constant-coefficient Lanchester equation. 2 Breakdown of Aggregate State-Space Models It has long been a reasonable hypothesis—but only that—that a relatively aggregated close battle in a particular area will have attrition that can be reason- 1 The principal reference for discussion of Lanchester equations is Taylor (1983b), which also covers many generalizations of the original work (Lanchester, 1916), including generalizations such as Bonder-Farrell theory (Bonder and Farrell, 1970) used in simulations. See also the recent collection of papers in Bracken et al. (1995), which includes historical analysis, a translation by Helmbold and Rehm of work by Osipov, and considerable thoughtful discussion. Wise (1991) explains some of the fundamental ambiguities in using and calibrating Lanchester laws. Hughes (1986) and Deitchman (1962) discuss applications of Lanchester models to naval and guerrilla warfare, respectively. Dupuy (1987) includes discussion of how his extensive history-based work on combat modeling relates (and does not relate) to Lanchester theory. 2 In fact, simulations do sometimes generate behaviors that look remarkably like what could be generated by such an equation, but that is an artifact of the particular application.
OCR for page 226
Technology for the United States Navy and Marine Corps, 2000-2035: Becoming a 21st-Century Force ably approximated by a state-space equation, that is, an equation relating the change in the sides' strengths (attrition) to various state variables and the duration of the time step, using the initial state-variable values of the time step and treating the combatants as all “in” the same location. The hypothesis clearly breaks down at low level (e.g., when evaluating alternative weapon systems in engagement-level combat where configural effects can be dominant (see also Appendix J )). The validity of the hypothesis also depends on there being many discrete countervailing microscopic processes (concentration and counterconcentration, ambush and withdrawal, fire and counterfire, and so on) that, over the time step and over the many replicas of the close battle across a theater, average to something relatively simple. This aggregate result may or may not correspond to a Lanchester square law, linear law, or something similar. It may be better described by the more general Bonder-Farrell equations, for example, but in some instances, there will be no such simplification because one side or the other has an asymmetric advantage that can be exploited because of multimodal probabilistic effects. The issue, then, becomes where and when various aggregate state-space models provide a good approximation of aggregate-level phenomena. It is inappropriate to draw broad conclusions, because contextual details matter a great deal to whether and which aggregations make sense. Myopia Caused by Head-on-Head Attrition While accusing simulation models of being Lanchesterian is often misleading, what critics who refer derisively to Lanchester models actually have in mind (clearly or dimly) is often something else, that most of today's theater-level models were designed from a so-called attrition perspective that conveys an image of war as mere head-on-head ground-force encounters with the two sides fighting to the bitter end. That is in contrast with a maneuver perspective in which campaigns consist of the sides maneuvering their forces in an attempt to create favorable circumstances of battle and to extricate themselves from unfavorable circumstances. Sometimes, a maneuver strategy can achieve victory without an extended attrition battle because one of the sides finds itself hopelessly outpositioned—and perhaps weakened by loss of critical assets or a collapse of command and control and unit coherence (possible objectives of information warfare). 3 Unfortunately, strategy and maneuver are often simplistic in models and studies conducted with the perspective of head-on-head attrition. Skilled users of even old-fashioned piston models can represent many effects of operational maneuver, but in practice, the result is often less impressive. 4 3 For discussion of information-warfare effects on theater combat, see Bonder et al. (1994). 4 For historical-empirical discussion of why standard attrition-warfare models are inadequate, see Rowland et al. (1996). Application to Korean analysis is described in Bennett (1995).
OCR for page 226
Technology for the United States Navy and Marine Corps, 2000-2035: Becoming a 21st-Century Force Theater- and Operational-Level Models Emphasizing Maneuver Warfare Although the head-on-head attrition modeling perspective is common, some theater-level models over the years have been designed to represent maneuver explicitly. For example, the IDAHEX model (Olsen, 1976) used in the 1980s introduced hexes and reintroduced interactive gaming with human players to make operational-level decisions; it did so specifically for the purpose of focusing effort on maneuver. 5 The Army War Colleges used a simpler but roughly comparable model in the early 1980s for similar reasons (the MTM). In the 1980s the RAND Strategy Assessment System (RSAS) was designed to focus attention on the strategy variable, introducing it explicitly in analytical war plans that included contingent branches and other adaptations. The RSAS also facilitated examining the consequences of nonattrition factors such as operational surprise, strategic flanking operations (e.g., Soviet use of the Austrian corridor), qualitative shortcomings in the fighting performance of some forces, the dependence of reserve-force effectiveness on training time before force employment (and the type of employment required of them), and the likely slowing effects of interdiction attacks. One version of the RSAS included a network model to improve the representation of flanking attacks, noncontiguous axes of advance, and critical nodes. 6 An improved version of the network model is incorporated in the JICM model, which has been used for extensive study of warfare in Korea, including warfare involving counteroffensives, flanking attacks, and asymmetric strategies involving weapons of mass destruction. 7 Another maneuver-oriented model was RAND's TLC/NLC, which was developed to a prototype stage using object-oriented programming and advanced graphics (Hillestad and Moore, 1996). Among other features, it included a rich network structure and reflected the Soviet correlation-of-force methodology for planning operational maneuver. While none of these models has been fully successful, while all of them share the severe shortcomings discussed below, and while even these maneuver-oriented models have sometimes been used in ways that reduce war to something looking like simplistic attrition warfare, the existence of the models and some of 5 It is of interest to note that the developer, Paul Olsen of the Institute of Defense Analyses, was criticized at the time (1986) because IDAHEX was not a “closed” model and, therefore, was allegedly inappropriate for analysis. His view was that without representing maneuver, the various more popular closed models were inappropriate. In fact, IDAHEX was later used extensively for analysis by the SHAPE Technical Center and a few other organizations. 6 For discussion of the RSAS, see Davis and Howe (1990), Bennett et al. (1992), and references therein. The network representation was due to earlier work by Patrick Allen and Barry Wilson. Some subsequent but unpublished documentation on RSAS 5.2 is also available through Bruce Bennett or Daniel Fox of RAND. 7 The JICM model is documented in Bennett (1994) and subsequent unpublished materials. It is used at RAND, the Army War College, OSD's PA&E, and some other organizations.
OCR for page 226
Technology for the United States Navy and Marine Corps, 2000-2035: Becoming a 21st-Century Force the studies accomplished with them demonstrates that the state of the art in combat modeling is substantially more advanced than those who decry head-on-head attrition modeling and Lanchester equations sometimes suggest. It is also significant to note that even older piston-style models have been used creatively and realistically— not only in research studies, but also by the operational commands, including the U.S. Central Command when preparing for Desert Storm. 8 Severe Limitations of Current Theater and Operational-Level Models While many criticisms of current models are exaggerated or overgeneralized, there is consensus throughout the community that DOD's current theater-and operational-level models are severely flawed. The major problems include their being overaggregated; having primitive or no representation of C4ISR, command-and-control, and information warfare; being almost exclusively deterministic; having too little representation of operational concepts, plans, and command; and having little ability to characterize the fluid and highly nonlinear combat operations anticipated for the future. Even many of the advanced features described above in connection with the RSAS and TLC efforts (notably those associated with decision models) no longer exist in operating models. Indeed, much of the current work with higher-level models such as TACWAR depends unreasonably on scripted representations of force employment, which are very difficult to work with because of the need for repeated iterations and tuning, and the absence of sufficiently adaptive behaviors. 9 WHERE NEXT? Research Opportunities for Improving Higher-Level Models Ideally, aggregate models should be informed by and even derived from more microscopic theory and experiment, including simulation “experiments” conducted at high resolution. Such experiments have their own shortcomings, but can nonetheless be a rich source of insight. 10 Furthermore, they are now 8 See articles by J.A. Appleget and F.T. Case et al., in Bracken et al. (1995). 9 Users are quite aware of these problems, of course. In the recent Deep-Attack Weapons-Mix Study (DAWMS), the Institute for Defense Analyses used a linear program (WORMS) to ensure that allocations of deep-attack weapons in TACWAR would be in some sense “optimal. ” Other aspects of the simulated campaign, however, were much less adaptive. 10 It does not follow that aggregate models must be derived from high-resolution models. Nor does it follow that some aggregate expression such as a Bonder-Farrel model or a Lanchester-square model, used locally, is invalid because some of those viewing the equations fail to see features they know are important microscopically (e.g., stochastic features, configural effects, and so on). More over, the usual claim that the validity of Lanchester equations depends on assumptions of homogeneous static forces with perfect local command and control is fallacious—the result of the classic blunder of confusing sufficient and necessary conditions. Lanchester equations are often motivated by simplistic models of combat, but the equations may be valid as aggregate-level descriptions in combat circumstances having none of the simplistic features said to be assumed. An analog here is that the ideal gas law does not depend on there being elastic billiard balls flying back and forth horizontally between walls (the “model” often used to motivate the law in high school chemistry). On the other hand, such aggregate depictions are clearly not valid in other cases and it is not currently clear when they are or are not.
OCR for page 226
Technology for the United States Navy and Marine Corps, 2000-2035: Becoming a 21st-Century Force feasible as the result of advances in computer science and entity-level simulation. As discussed in the text, exploiting this opportunity should be given high priority. While there will continue to be an important role for diverse state-space models implemented as simulations—including some that will continue to be incorrectly characterized as “Lanchester models”—a great deal of effort is needed to establish a better foundation for the assumptions used in those models. For example, many of the terrain factors used in combat models were estimated many years ago when it was computationally impossible to conduct high-resolution high-quality simulations in the numbers needed to identify good aggregate representations. So also, the assumption of deterministic aggregate behavior was made in part because it was computationally infeasible to do otherwise. That multimodal distributions aggregate into something simpler, which can be treated by deterministic equations (plus uncertainty analysis to account for important branches) was and is a reasonable hypothesis for higher-level battle, but we do not currently know when the hypothesis is correct. It is clear that it fails for engagements in which one side can consistently exploit a range advantage. In any case, a new round of such research is now possible and needed. While current DOD simulations are in some cases based on earlier research that included comparisons with higher-resolution simulations (see, e.g., Farrell (1989), which discusses the early development of Vector models), that work should be reopened since the quality of the high-resolution simulations is now so much better. In pursuing a research program to connect the worlds of high- and low-resolution modeling, it is essential to recognize information resides at all levels of aggregation and to avoid a pure bottom-up approach. Instead, the ideal is an approach in which models of differing resolution are used to exploit all the information available and, then, to cross-calibrate each other. Constructing such mutually calibrated families of models is a major undertaking, as discussed in Chapter 6 and Appendix E . In conclusion, while we have attempted to illuminate issues regarding Lanchester models and DOD simulations, and to soften exaggerations, we emphasize the need for in-depth research to better understand the phenomenology of combat. One important spinoff of this could be DOD's emerging with integrated or semi-integrated families of mutually calibrated models appropriate for the full
OCR for page 226
Technology for the United States Navy and Marine Corps, 2000-2035: Becoming a 21st-Century Force range of M&S applications. The traditional approach of developing models separately for the various levels of resolution is fatally flawed when it means working with blinders on, which it often does. Those working exclusively at low resolutions are unlikely to understand the underlying phenomena and are therefore likely to misrepresent the aggregate phenomena. Those working exclusively at high resolution are unlikely to understand larger contexts and interrelationships. Their insights and conclusions may be much more conditionally valid than they realize. Further, the calibration of high-resolution models should exploit all the relevant information available, much of which is at low resolution. We conclude, then, that research should be conducted jointly at multiple levels of resolution with a great deal of interaction and the goal of integration. 11 Such work should include relating stochastic and deterministic representations, as well as consideration of many other types of uncertainty. Such things will not occur without changes in both funding and management practices. The Department of the Navy should advocate such changes strenuously in the joint arena and with OSD. Otherwise, it is likely that the next generation of aggregate combat models for use in joint analysis will not be significantly better than the ones already available —and perhaps worse. 11 This integration need not be in any single model, however. It might be in textbooks plus occasional cross-calibrations. We are not advocating single do-it-all-comprehensively models.