Canadian soldiers going “over the top” during an assault in the First World War. [History.com][This post was originally published on 1 December 2017.]
How many troops are needed to successfully attack or defend on the battlefield? There is a long-standing rule of thumb that holds that an attacker requires a 3-1 preponderance over a defender in combat in order to win. The aphorism is so widely accepted that few have questioned whether it is actually true or not.
Trevor Dupuy challenged the validity of the 3-1 rule on empirical grounds. He could find no historical substantiation to support it. In fact, his research on the question of force ratios suggested that there was a limit to the value of numerical preponderance on the battlefield.
TDI President Chris Lawrence has also challenged the 3-1 rule in his own work on the subject.
The validity of the 3-1 rule is no mere academic question. It underpins a great deal of U.S. military policy and warfighting doctrine. Yet, the only time the matter was seriously debated was in the 1980s with reference to the problem of defending Western Europe against the threat of Soviet military invasion.
It is probably long past due to seriously challenge the validity and usefulness of the 3-1 rule again.
Battle of Chancellorsville by Kurz and Allison (1888). This painting depicted the wounding of Confederate General Thomas J. “Stonewall” Jackson on 2 May 1863, while leading one of the more famous flank attacks in history.
Flank or rear attack is more likely to succeed than frontal attack. Among the many reasons for this are the following: there is greater opportunity for surprise by the attacker; the defender cannot be strong everywhere at once, and the front is the easiest focus for defensive effort; and the morale of the defender tends to be shaken when the danger of encirclement is evident. Again, historical examples are numerous, beginning with Hannibal’s tactical plans and brilliant executions of the Battles of Lake Trasimene and Cannae. Any impression that the concept of envelopment or of a “strategy of indirect approach” has arisen either from the introduction of modern weapons of war, or from the ruminations of recent writers on military affairs, is a grave misperception of history and underestimates earlier military thinkers.
“Seek the flanks” has been a military adage since antiquity, but its significance was enhanced tremendously when the conoidal bullet of the breech-loading, rifled musket revolutionized warfare in the mid-nineteenth century. This led Moltke to his 1867 observation that the increased deadliness of firepower demanded that the strategic offensive be coupled with tactical defensive, an idea that depended upon strategic envelopment for its accomplishment. This was a basic element of Moltke‘s strategy in the 1870 campaign in France. Its tactical manifestations took place at Metz and Sedan; both instances in which the Germans took up defensive positions across the French line of communications to Paris, and the French commanders, forced to attack, were defeated.
The essential emphasis of modern tactics and operational art remains enabling flank or rear attacks on enemy forces in order to obtain decisive results in combat. Will this remain true in the future? The ongoing historical pattern of ground forces dispersing on the battlefield in response to the increasing lethality of weapons seems likely to enhance the steadily increasing non-linear and non-contiguous character of modern battles in both conventional and irregular warfare.
The architects of the U.S. multi-domain battle and operations doctrine seem to anticipate this. Highly dispersed and distributed future battlefields are likely to offer constant, multiple opportunities (and risks) for flank and rear attacks. They are also likely to scramble current efforts to shape and map future battlefield geometry and architecture.
Even though offensive action is essential to ultimate combat success, a combat commander opposed by a more powerful enemy has no choice but to assume a defensive posture. Since defensive posture automatically increases the combat power of his force, the defending commander at least partially redresses the imbalance of forces. At a minimum he is able to slow down the advance of the attacking enemy, and he might even beat him. In this way, through negative combat results, the defender may ultimately hope to wear down the attacker to the extent that his initial relative weakness is transformed into relative superiority, thus offering the possibility of eventually assuming the offensive and achieving positive combat results. The Franklin and Nashville Campaign of our Civil War, and the El Alamein Campaign of World War II are examples.
Sometimes the commander of a numerically superior offensive force may reduce the strength of portions of his force in order to achieve decisive superiority for maximum impact on the enemy at some other critical point on the battlefield, with the result that those reduced-strength components are locally outnumbered. A contingent thus reduced in strength may therefore be required to assume a defensive posture, even though the overall operational posture of the marginally superior force is offensive, and the strengthened contingent of the same force is attacking with the advantage of superior combat power. A classic example was the role of Davout at Auerstadt when Napoléon was crushing the Prussians at Jena. Another is the role played by “Stonewall” Jackson’s corps at the Second Battle of Bull Run. [pp. 2-3]
This verity is both derivative of Dupuy’s belief that the defensive posture is a human reaction to the lethal environment of combat, and his concurrence with Clausewitz’s dictum that the defense is the stronger form of combat. Soldiers in combat will sometimes reach a collective conclusion that they can no longer advance in the face of lethal opposition, and will stop and seek cover and concealment to leverage the power of the defense. Exploiting the multiplying effect of the defensive is also a way for a force with weaker combat power to successfully engage a stronger one.
Minimum essential means must be employed at points other than that of decision. To devote means to unnecessary secondary efforts or to employ excessive means on required secondary efforts is to violate the principle of both mass and the objective. Limited attacks, the defensive, deception, or even retrograde action are used in noncritical areas to achieve mass in the critical area.
These concepts are well ingrained in modern U.S. Army doctrine. FM 3-0 Operations (2017) summarizes the defensive this way:
Defensive tasks are conducted to defeat an enemy attack, gain time, economize forces, and develop conditions favorable for offensive or stability tasks. Normally, the defense alone cannot achieve a decisive victory. However, it can set conditions for a counteroffensive or counterattack that enables Army forces to regain and exploit the initiative. Defensive tasks are a counter to enemy offensive actions. They defeat attacks, destroying as much of an attacking enemy as possible. They also preserve and maintain control over land, resources, and populations. The purpose of defensive tasks is to retain key terrain, guard populations, protect lines of communications, and protect critical capabilities against enemy attacks and counterattacks. Commanders can conduct defensive tasks to gain time and economize forces, so offensive tasks can be executed elsewhere. [Para 1-72]
UPDATE: Just as I posted this, out comes a contrarian view from U.S. Army CAPT Brandon Morgan via the Modern War Institute at West Point blog. He argues that the U.S. Army is not placing enough emphasis on preparing to conduct defensive operations:
In his seminal work On War, Carl von Clausewitz famously declared that, in comparison to the offense, “the defensive form of warfare is intrinsically stronger than the offensive.”
This is largely due to the defender’s ability to occupy key terrain before the attack, and is most true when there is sufficient time to prepare the defense. And yet within the doctrinal hierarchy of the four elements of decisive action (offense, defense, stability, and defense support of civil authorities), the US Army prioritizes offensive operations. Ultimately, this has led to training that focuses almost exclusively on offensive operations at the cost of deliberate planning for the defense. But in the context of a combined arms fight against a near-peer adversary, US Army forces will almost assuredly find themselves initially fighting in a defense. Our current neglect of deliberate planning for the defense puts these soldiers who will fight in that defense at grave risk.
[Prussian military theorist, Carl von] Clausewitz expressed this: “Defense is the stronger form of combat.” It is possible to demonstrate by the qualitative comparison of many battles that Clausewitz is right and that posture has a multiplicative effect on the combat power of a military force that takes advantage of terrain and fortifications, whether hasty and rudimentary, or intricate and carefully prepared. There are many well-known examples of the need of an attacker for a preponderance of strength in order to carry the day against a well-placed and fortified defender. One has only to recall Thermopylae, the Alamo, Fredericksburg, Petersburg, and El Alamein to realize the advantage enjoyed by a defender with smaller forces, well placed, and well protected. [p. 2]
The advantages of fighting on the defensive and the benefits of cover and concealment in certain types of terrain have long been basic tenets in military thinking. Dupuy, however, considered defensive combat posture and defensive value of terrain not just to be additive, but combat power multipliers, or circumstantial variables of combat that when skillfully applied and exploited, the effects of which could increase the overall fighting capability of a military force.
The statement [that the defensive is the stronger form of combat] implies a comparison of relative strength. It is essentially scalar and thus ultimately quantitative. Clausewitz did not attempt to define the scale of his comparison. However, by following his conceptual approach it is possible to establish quantities for this comparison. Depending upon the extent to which the defender has had the time and capability to prepare for defensive combat, and depending also upon such considerations as the nature of the terrain which he is able to utilize for defense, my research tells me that the comparative strength of defense to offense can range from a factor with a minimum value of about 1.3 to maximum value of more than 3.0. [p. 26]
The values Dupuy established for posture and terrain based on historical combat experience were as follows:
For example, Dupuy calculated that mounting even a hasty defense in rolling, gentle terrain with some vegetation could increase a force’s combat power by more than 50%. This is a powerful effect, achievable without the addition of any extra combat capability.
It should be noted that these values are both descriptive, in terms of defining Dupuy’s theoretical conception of the circumstantial variables of combat, as well as factors specifically calculated for use in his combat models. Some of these factors have found their way into models and simulations produced by others and some U.S. military doctrinal publications, usually without attribution and shorn of explanatory context. (A good exploration of the relationship between the values Dupuy established for the circumstantial variables of combat and his combat models, and the pitfalls of applying them out of context can be found here.)
While the impact of terrain on combat is certainly an integral part of current U.S. Army thinking at all levels, and is constantly factored into combat planning and assessment, its doctrine does not explicitly acknowledge the classic Clausewitzian notion of a power disparity between the offense and defense. Nor are the effects of posture or terrain thought of as combat multipliers.
However, the Army does implicitly recognize the advantage of the defensive through its stubbornly persistent adherence to the so-called 3-1 rule of combat. Its version of this (which the U.S. Marine Corps also uses) is described in doctrinal publications as “historical minimum planning ratios,” which proscribe that a 3-1 advantage in numerical force ratio is necessary for an attacker to defeat a defender in a prepared or fortified position. Overcoming a defender in a hasty defense posture requires a 2.5-1 force ratio advantage. The force ratio advantages the Army considers necessary for decisive operations are even higher. While the 3-1 rule is a deeply problematic construct, the fact that is the only quantitative planning factor included in current doctrine reveals a healthy respect for the inherent power of the defensive.
Response to Niklas Zetterling’s Article by Christopher A. Lawrence
Mr. Zetterling is currently a professor at the Swedish War College and previously worked at the Swedish National Defense Research Establishment. As I have been having an ongoing dialogue with Prof. Zetterling on the Battle of Kursk, I have had the opportunity to witness his approach to researching historical data and the depth of research. I would recommend that all of our readers take a look at his recent article in the Journal of Slavic Military Studies entitled “Loss Rates on the Eastern Front during World War II.” Mr. Zetterling does his German research directly from the Captured German Military Records by purchasing the rolls of microfilm from the US National Archives. He is using the same German data sources that we are. Let me attempt to address his comments section by section:
The Database on Italy 1943-44:
Unfortunately, the Italian combat data was one of the early HERO research projects, with the results first published in 1971. I do not know who worked on it nor the specifics of how it was done. There are references to the Captured German Records, but significantly, they only reference division files for these battles. While I have not had the time to review Prof. Zetterling‘s review of the original research. I do know that some of our researchers have complained about parts of the Italian data. From what I’ve seen, it looks like the original HERO researchers didn’t look into the Corps and Army files, and assumed what the attached Corps artillery strengths were. Sloppy research is embarrassing, although it does occur, especially when working under severe financial constraints (for example, our Battalion-level Operations Database). If the research is sloppy or hurried, or done from secondary sources, then hopefully the errors are random, and will effectively counterbalance each other, and not change the results of the analysis. If the errors are all in one direction, then this will produce a biased result.
I have no basis to believe that Prof. Zetterling’s criticism is wrong, and do have many reasons to believe that it is correct. Until l can take the time to go through the Corps and Army files, I intend to operate under the assumption that Prof. Zetterling’s corrections are good. At some point I will need to go back through the Italian Campaign data and correct it and update the Land Warfare Database. I did compare Prof. Zetterling‘s list of battles with what was declared to be the forces involved in the battle (according to the Combat Data Subscription Service) and they show the following attached artillery:
It is clear that the battles were based on the assumption that here was Corps-level German artillery. A strength comparison between the two sides is displayed in the chart on the next page.
The Result Formula:
CEV is calculated from three factors. Therefore a consistent 20% error in casualties will result in something less than a 20% error in CEV. The mission effectiveness factor is indeed very “fuzzy,” and these is simply no systematic method or guidance in its application. Sometimes, it is not based upon the assigned mission of the unit, but its perceived mission based upon the analyst’s interpretation. But, while l have the same problems with the mission accomplishment scores as Mr. Zetterling, I do not have a good replacement. Considering the nature of warfare, I would hate to create CEVs without it. Of course, Trevor Dupuy was experimenting with creating CEVs just from casualty effectiveness, and by averaging his two CEV scores (CEVt and CEVI) he heavily weighted the CEV calculation for the TNDM towards measuring primarily casualty effectiveness (see the article in issue 5 of the Newsletter, “Numerical Adjustment of CEV Results: Averages and Means“). At this point, I would like to produce a new, single formula for CEV to replace the current two and its averaging methodology. I am open to suggestions for this.
Supply Situation:
The different ammunition usage rate of the German and US Armies is one of the reasons why adding a logistics module is high on my list of model corrections. This was discussed in Issue 2 of the Newsletter, “Developing a Logistics Model for the TNDM.” As Mr. Zetterling points out, “It is unlikely that an increase in artillery ammunition expenditure will result in a proportional increase in combat power. Rather it is more likely that there is some kind of diminished return with increased expenditure.” This parallels what l expressed in point 12 of that article: “It is suspected that this increase [in OLIs] will not be linear.”
The CEV does include “logistics.” So in effect, if one had a good logistics module, the difference in logistics would be accounted for, and the Germans (after logistics is taken into account) may indeed have a higher CEV.
General Problems with Non-Divisional Units Tooth-to-Tail Ratio
Point taken. The engagements used to test the TNDM have been gathered over a period of over 25 years, by different researchers and controlled by different management. What is counted when and where does change from one group of engagements to the next. While l do think this has not had a significant result on the model outcomes, it is “sloppy” and needs to be addressed.
The Effects of Defensive Posture
This is a very good point. If the budget was available, my first step in “redesigning” the TNDM would be to try to measure the effects of terrain on combat through the use of a large LWDB-type database and regression analysis. I have always felt that with enough engagements, one could produce reliable values for these figures based upon something other than judgement. Prof. Zetterling’s proposed methodology is also a good approach, easier to do, and more likely to get a conclusive result. I intend to add this to my list of model improvements.
Conclusions
There is one other problem with the Italian data that Prof. Zetterling did not address. This was that the Germans and the Allies had different reporting systems for casualties. Quite simply, the Germans did not report as casualties those people who were lightly wounded and treated and returned to duty from the divisional aid station. The United States and England did. This shows up when one compares the wounded to killed ratios of the various armies, with the Germans usually having in the range of 3 to 4 wounded for every one killed, while the allies tend to have 4 to 5 wounded for every one killed. Basically, when comparing the two reports, the Germans “undercount” their casualties by around 17 to 20%. Therefore, one probably needs to use a multiplier of 20 to 25% to match the two casualty systems. This was not taken into account in any the work HERO did.
Because Trevor Dupuy used three factors for measuring his CEV, this error certainly resulted in a slightly higher CEV for the Germans than should have been the case, but not a 20% increase. As Prof. Zetterling points out, the correction of the count of artillery pieces should result in a higher CEV than Col. Dupuy calculated. Finally, if Col. Dupuy overrated the value of defensive terrain, then this may result in the German CEV being slightly lower.
As you may have noted in my list of improvements (Issue 2, “Planned Improvements to the TNDM”), I did list “revalidating” to the QJM Database. [NOTE: a summary of the QJM/TNDM validation efforts can be found here.] As part of that revalidation process, we would need to review the data used in the validation data base first, account for the casualty differences in the reporting systems, and determine if the model indeed overrates the effect of terrain on defense.
Perhaps one of the most debated results of the TNDM (and its predecessors) is the conclusion that the German ground forces on average enjoyed a measurable qualitative superiority over its US and British opponents. This was largely the result of calculations on situations in Italy in 1943-44, even though further engagements have been added since the results were first presented. The calculated German superiority over the Red Army, despite the much smaller number of engagements, has not aroused as much opposition. Similarly, the calculated Israeli effectiveness superiority over its enemies seems to have surprised few.
However, there are objections to the calculations on the engagements in Italy 1943. These concern primarily the database, but there are also some questions to be raised against the way some of the calculations have been made, which may possibly have consequences for the TNDM.
Here it is suggested that the German CEV [combat effectiveness value] superiority was higher than originally calculated. There are a number of flaws in the original calculations, each of which will be discussed separately below. With the exception of one issue, all of them, if corrected, tend to give a higher German CEV.
The Database on Italy 1943-44
According to the database the German divisions had considerable fire support from GHQ artillery units. This is the only possible conclusion from the fact that several pieces of the types 15cm gun, 17cm gun, 21cm gun, and 15cm and 21cm Nebelwerfer are included in the data for individual engagements. These types of guns were almost exclusively confined to GHQ units. An example from the database are the three engagements Port of Salerno, Amphitheater, and Sele-Calore Corridor. These take place simultaneously (9-11 September 1943) with the German 16th Pz Div on the Axis side in all of them (no other division is included in the battles). Judging from the manpower figures, it seems to have been assumed that the division participated with one quarter of its strength in each of the two former battles and half its strength in the latter. According to the database, the number of guns were:
15cm gun
28
17cm gun
12
21cm gun
12
15cm NbW
27
21cm NbW
21
This would indicate that the 16th Pz Div was supported by the equivalent of more than five non-divisional artillery battalions. For the German army this is a suspiciously high number, usually there were rather something like one GHQ artillery battalion for each division, or even less. Research in the German Military Archives confirmed that the number of GHQ artillery units was far less than indicated in the HERO database. Among the useful documents found were a map showing the dispositions of 10th Army artillery units. This showed clearly that there was only one non-divisional artillery unit south of Rome at the time of the Salerno landings, the III/71 Nebelwerfer Battalion. Also the 557th Artillery Battalion (17cm gun) was present, it was included in the artillery regiment (33rd Artillery Regiment) of 15th Panzergrenadier Division during the second half of 1943. Thus the number of German artillery pieces in these engagements is exaggerated to an extent that cannot be considered insignificant. Since OLI values for artillery usually constitute a significant share of the total OLI of a force in the TNDM, errors in artillery strength cannot be dismissed easily.
While the example above is but one, further archival research has shown that the same kind of error occurs in all the engagements in September and October 1943. It has not been possible to check the engagements later during 1943, but a pattern can be recognized. The ratio between the numbers of various types of GHQ artillery pieces does not change much from battle to battle. It seems that when the database was developed, the researchers worked with the assumption that the German corps and army organizations had organic artillery, and this assumption may have been used as a “rule of thumb.” This is wrong, however; only artillery staffs, command and control units were included in the corps and army organizations, not firing units. Consequently we have a systematic error, which cannot be corrected without changing the contents of the database. It is worth emphasizing that we are discussing an exaggeration of German artillery strength of about 100%, which certainly is significant. Comparing the available archival records with the database also reveals errors in numbers of tanks and antitank guns, but these are much smaller than the errors in artillery strength. Again these errors do always inflate the German strength in those engagements l have been able to check against archival records. These errors tend to inflate German numerical strength, which of course affects CEV calculations. But there are further objections to the CEV calculations.
The Result Formula
The “result formula” weighs together three factors: casualties inflicted, distance advanced, and mission accomplishment. It seems that the first two do not raise many objections, even though the relative weight of them may always be subject to argumentation.
The third factor, mission accomplishment, is more dubious however. At first glance it may seem to be natural to include such a factor. Alter all, a combat unit is supposed to accomplish the missions given to it. However, whether a unit accomplishes its mission or not depends both on its own qualities as well as the realism of the mission assigned. Thus the mission accomplishment factor may reflect the qualities of the combat unit as well as the higher HQs and the general strategic situation. As an example, the Rapido crossing by the U.S. 36th Infantry Division can serve. The division did not accomplish its mission, but whether the mission was realistic, given the circumstances, is dubious. Similarly many German units did probably, in many situations, receive unrealistic missions, particularly during the last two years of the war (when most of the engagements in the database were fought). A more extreme example of situations in which unrealistic missions were given is the battle in Belorussia, June-July 1944, where German units were regularly given impossible missions. Possibly it is a general trend that the side which is fighting at a strategic disadvantage is more prone to give its combat units unrealistic missions.
On the other hand it is quite clear that the mission assigned may well affect both the casualty rates and advance rates. If, for example, the defender has a withdrawal mission, advance may become higher than if the mission was to defend resolutely. This must however not necessarily be handled by including a missions factor in a result formula.
I have made some tentative runs with the TNDM, testing with various CEV values to see which value produced an outcome in terms of casualties and ground gained as near as possible to the historical result. The results of these runs are very preliminary, but the tendency is that higher German CEVs produce more historical outcomes, particularly concerning combat.
Supply Situation
According to scattered information available in published literature, the U.S. artillery fired more shells per day per gun than did German artillery. In Normandy, US 155mm M1 howitzers fired 28.4 rounds per day during July, while August showed slightly lower consumption, 18 rounds per day. For the 105mm M2 howitzer the corresponding figures were 40.8 and 27.4. This can be compared to a German OKH study which, based on the experiences in Russia 1941-43, suggested that consumption of 105mm howitzer ammunition was about 13-22 rounds per gun per day, depending on the strength of the opposition encountered. For the 150mm howitzer the figures were 12-15.
While these figures should not be taken too seriously, as they are not from primary sources and they do also reflect the conditions in different theaters, they do at least indicate that it cannot be taken for granted that ammunition expenditure is proportional to the number of gun barrels. In fact there also exist further indications that Allied ammunition expenditure was greater than the German. Several German reports from Normandy indicate that they were astonished by the Allied ammunition expenditure.
It is unlikely that an increase in artillery ammunition expenditure will result in a proportional increase combat power. Rather it is more likely that there is some kind of diminished return with increased expenditure.
General Problems with Non-Divisional Units
A division usually (but not necessarily) includes various support services, such as maintenance, supply, and medical services. Non-divisional combat units have to a greater extent to rely on corps and army for such support. This makes it complicated to include such units, since when entering, for example, the manpower strength and truck strength in the TNDM, it is difficult to assess their contribution to the overall numbers.
Furthermore, the amount of such forces is not equal on the German and Allied sides. In general the Allied divisional slice was far greater than the German. In Normandy the US forces on 25 July 1944 had 812,000 men on the Continent, while the number of divisions was 18 (including the 5th Armored, which was in the process of landing on the 25th). This gives a divisional slice of 45,000 men. By comparison the German 7th Army mustered 16 divisions and 231,000 men on 1 June 1944, giving a slice of 14,437 men per division. The main explanation for the difference is the non-divisional combat units and the logistical organization to support them. In general, non-divisional combat units are composed of powerful, but supply-consuming, types like armor, artillery, antitank and antiaircraft. Thus their contribution to combat power and strain on the logistical apparatus is considerable. However I do not believe that the supporting units’ manpower and vehicles have been included in TNDM calculations.
There are however further problems with non-divisional units. While the whereabouts of tank and tank destroyer units can usually be established with sufficient certainty, artillery can be much harder to pin down to a specific division engagement. This is of course a greater problem when the geographical extent of a battle is small.
Tooth-to-Tail Ratio
Above was discussed the lack of support units in non-divisional combat units. One effect of this is to create a force with more OLI per man. This is the result of the unit‘s “tail” belonging to some other part of the military organization.
In the TNDM there is a mobility formula, which tends to favor units with many weapons and vehicles compared to the number of men. This became apparent when I was performing a great number of TNDM runs on engagements between Swedish brigades and Soviet regiments. The Soviet regiments usually contained rather few men, but still had many AFVs, artillery tubes, AT weapons, etc. The Mobility Formula in TNDM favors such units. However, I do not think this reflects any phenomenon in the real world. The Soviet penchant for lean combat units, with supply, maintenance, and other services provided by higher echelons, is not a more effective solution in general, but perhaps better suited to the particular constraints they were experiencing when forming units, training men, etc. In effect these services were existing in the Soviet army too, but formally not with the combat units.
This problem is to some extent reminiscent to how density is calculated (a problem discussed by Chris Lawrence in a recent issue of the Newsletter). It is comparatively easy to define the frontal limit of the deployment area of force, and it is relatively easy to define the lateral limits too. It is, however, much more difficult to say where the rear limit of a force is located.
When entering forces in the TNDM a rear limit is, perhaps unintentionally, drawn. But if the combat unit includes support units, the rear limit is pushed farther back compared to a force whose combat units are well separated from support units.
To what extent this affects the CEV calculations is unclear. Using the original database values, the German forces are perhaps given too high combat strength when the great number of GHQ artillery units is included. On the other hand, if the GHQ artillery units are not included, the opposite may be true.
The Effects of Defensive Posture
The posture factors are difficult to analyze, since they alone do not portray the advantages of defensive position. Such effects are also included in terrain factors.
It seems that the numerical values for these factors were assigned on the basis of professional judgement. However, when the QJM was developed, it seems that the developers did not assume the German CEV superiority. Rather, the German CEV superiority seems to have been discovered later. It is possible that the professional judgement was about as wrong on the issue of posture effects as they were on CEV. Since the British and American forces were predominantly on the offensive, while the Germans mainly defended themselves, a German CEV superiority may, at least partly, be hidden in two high effects for defensive posture.
When using corrected input data on the 20 situations in Italy September-October 1943, there is a tendency that the German CEV is higher when they attack. Such a tendency is also discernible in the engagements presented in Hitler’s Last Gamble. Appendix H, even though the number of engagements in the latter case is very small.
As it stands now this is not really more than a hypothesis, since it will take an analysis of a greater number of engagements to confirm it. However, if such an analysis is done, it must be done using several sets of data. German and Allied attacks must be analyzed separately, and preferably the data would be separated further into sets for each relevant terrain type. Since the effects of the defensive posture are intertwined with terrain factors, it is very much possible that the factors may be correct for certain terrain types, while they are wrong for others. It may also be that the factors can be different for various opponents (due to differences in training, doctrine, etc.). It is also possible that the factors are different if the forces are predominantly composed of armor units or mainly of infantry.
One further problem with the effects of defensive position is that it is probably strongly affected by the density of forces. It is likely that the main effect of the density of forces is the inability to use effectively all the forces involved. Thus it may be that this factor will not influence the outcome except when the density is comparatively high. However, what can be regarded as “high” is probably much dependent on terrain, road net quality, and the cross-country mobility of the forces.
Conclusions
While the TNDM has been criticized here, it is also fitting to praise the model. The very fact that it can be criticized in this way is a testimony to its openness. In a sense a model is also a theory, and to use Popperian terminology, the TNDM is also very testable.
It should also be emphasized that the greatest errors are probably those in the database. As previously stated, I can only conclude safely that the data on the engagements in Italy in 1943 are wrong; later engagements have not yet been checked against archival documents. Overall the errors do not represent a dramatic change in the CEV values. Rather, the Germans seem to have (in Italy 1943) a superiority on the order of 1.4-1.5, compared to an original figure of 1.2-1.3.
During September and October 1943, almost all the German divisions in southern Italy were mechanized or parachute divisions. This may have contributed to a higher German CEV. Thus it is not certain that the conclusions arrived at here are valid for German forces in general, even though this factor should not be exaggerated, since many of the German divisions in Italy were either newly raised (e.g., 26th Panzer Division) or rebuilt after the Stalingrad disaster (16th Panzer Division plus 3rd and 29th Panzergrenadier Divisions) or the Tunisian debacle (15th Panzergrenadier Division).
This is like saying, “A team can’t score in football unless it has the ball.” Although subsequent verities stress the strength, value, and importance of defense, this should not obscure the essentiality of offensive action to ultimate combat success. Even in instances where a defensive strategy might conceivably assure a favorable war outcome—as was the case of the British against Napoleon, and as the Confederacy attempted in the American Civil War—selective employment of offensive tactics and operations is required if the strategic defender is to have any chance of final victory. [pp. 1-2]
Only offensive action achieves decisive results. Offensive action permits the commander to exploit the initiative and impose his will on the enemy. The defensive may be forced on the commander, but it should be deliberately adopted only as a temporary expedient while awaiting an opportunity for offensive action or for the purpose of economizing forces on a front where a decision is not sought. Even on the defensive the commander seeks every opportunity to seize the initiative and achieve decisive results by offensive action. [Original emphasis]
Interestingly enough, the offensive no longer retains its primary place in current Army doctrinal thought. The Army consigned its list of the principles of war to an appendix in the 2008 edition of FM 3-0 Operations and omitted them entirely from the 2017 revision. As the current edition of FM 3-0 Operations lays it out, the offensive is now placed on the same par as the defensive and stability operations:
Unified land operations are simultaneous offensive, defensive, and stability or defense support of civil authorities’ tasks to seize, retain, and exploit the initiative to shape the operational environment, prevent conflict, consolidate gains, and win our Nation’s wars as part of unified action (ADRP 3-0)…
At the heart of the Army’s operational concept is decisive action. Decisive action is the continuous, simultaneous combinations of offensive, defensive, and stability or defense support of civil authorities tasks (ADRP 3-0). During large-scale combat operations, commanders describe the combinations of offensive, defensive, and stability tasks in the concept of operations. As a single, unifying idea, decisive action provides direction for an entire operation. [p. I-16; original emphasis]
It is perhaps too easy to read too much into this change in emphasis. On the very next page, FM 3-0 describes offensive “tasks” thusly:
Offensive tasks are conducted to defeat and destroy enemy forces and seize terrain, resources, and population centers. Offensive tasks impose the commander’s will on the enemy. The offense is the most direct and sure means of seizing and exploiting the initiative to gain physical and cognitive advantages over an enemy. In the offense, the decisive operation is a sudden, shattering action that capitalizes on speed, surprise, and shock effect to achieve the operation’s purpose. If that operation does not destroy or defeat the enemy, operations continue until enemy forces disintegrate or retreat so they no longer pose a threat. Executing offensive tasks compels an enemy to react, creating or revealing additional weaknesses that an attacking force can exploit. [p. I-17]
The change in emphasis likely reflects recent U.S. military experience where decisive action has not yielded much in the way of decisive outcomes, as is mentioned in FM 3-0’s introduction. Joint force offensives in 2001 and 2003 “achieved rapid initial military success but no enduring political outcome, resulting in protracted counterinsurgency campaigns.” The Army now anticipates a future operating environment where joint forces can expect to “work together and with unified action partners to successfully prosecute operations short of conflict, prevail in large-scale combat operations, and consolidate gains to win enduring strategic outcomes” that are not necessarily predicated on offensive action alone. We may have to wait for the next edition of FM 3-0 to see if the Army has drawn valid conclusions from the recent past or not.
Consistent Scoring of Weapons and Aggregation of Forces: The Cornerstone of Dupuy’s Quantitative Analysis of Historical Land Battles by
James G. Taylor, PhD,
Dept. of Operations Research, Naval Postgraduate School
Introduction
Col. Trevor N. Dupuy was an American original, especially as regards the quantitative study of warfare. As with many prophets, he was not entirely appreciated in his own land, particularly its Military Operations Research (OR) community. However, after becoming rather familiar with the details of his mathematical modeling of ground combat based on historical data, I became aware of the basic scientific soundness of his approach. Unfortunately, his documentation of methodology was not always accepted by others, many of whom appeared to confuse lack of mathematical sophistication in his documentation with lack of scientific validity of his basic methodology.
The purpose of this brief paper is to review the salient points of Dupuy’s methodology from a system’s perspective, i.e., to view his methodology as a system, functioning as an organic whole to capture the essence of past combat experience (with an eye towards extrapolation into the future). The advantage of this perspective is that it immediately leads one to the conclusion that if one wants to use some functional relationship derived from Dupuy’s work, then one should use his methodologies for scoring weapons, aggregating forces, and adjusting for operational circumstances; since this consistency is the only guarantee of being able to reproduce historical results and to project them into the future.
Implications (of this system’s perspective on Dupuy’s work) for current DOD models will be discussed. In particular, the Military OR community has developed quantitative methods for imputing values to weapon systems based on their attrition capability against opposing forces and force interactions.[1] One such approach is the so-called antipotential-potential method[2] used in TACWAR[3] to score weapons. However, one should not expect such scores to provide valid casualty estimates when combined with historically derived functional relationships such as the so-called ATLAS casualty-rate curves[4] used in TACWAR, because a different “yard-stick” (i.e. measuring system for estimating the relative combat potential of opposing forces) was used to develop such a curve.
Overview of Dupuy’s Approach
This section briefly outlines the salient features of Dupuy’s approach to the quantitative analysis and modeling of ground combat as embodied in his Tactical Numerical Deterministic Model (TNDM) and its predecessor the Quantified Judgment Model (QJM). The interested reader can find details in Dupuy [1979] (see also Dupuy [1985][5], [1987], [1990]). Here we will view Dupuy’s methodology from a system approach, which seeks to discern its various components and their interactions and to view these components as an organic whole. Essentially Dupuy’s approach involves the development of functional relationships from historical combat data (see Fig. 1) and then using these functional relationships to model future combat (see Fig, 2).
At the heart of Dupuy’s method is the investigation of historical battles and comparing the relationship of inputs (as quantified by relative combat power, denoted as Pa/Pd for that of the attacker relative to that of the defender in Fig. l)(e.g. see Dupuy [1979, pp. 59-64]) to outputs (as quantified by extent of mission accomplishment, casualty effectiveness, and territorial effectiveness; see Fig. 2) (e.g. see Dupuy [1979, pp. 47-50]), The salient point is that within this scheme, the main input[6] (i.e. relative combat power) to a historical battle is a derived quantity. It is computed from formulas that involve three essential aspects: (1) the scoring of weapons (e.g, see Dupuy [1979, Chapter 2 and also Appendix A]), (2) aggregation methodology for a force (e.g. see Dupuy [1979, pp. 43-46 and 202-203]), and (3) situational-adjustment methodology for determining the relative combat power of opposing forces (e.g. see Dupuy [1979, pp. 46-47 and 203-204]). In the force-aggregation step the effects on weapons of Dupuy’s environmental variables and one operational variable (air superiority) are considered[7], while in the situation-adjustment step the effects on forces of his behavioral variables[8] (aggregated into a single factor called the relative combat effectiveness value (CEV)) and also the other operational variables are considered (Dupuy [1987, pp. 86-89])
Figure 1.
Moreover, any functional relationships developed by Dupuy depend (unless shown otherwise) on his computational system for derived quantities, namely OLls, force strengths, and relative combat power. Thus, Dupuy’s results depend in an essential manner on his overall computational system described immediately above. Consequently, any such functional relationship (e.g. casualty-rate curve) directly or indirectly derivative from Dupuy‘s work should still use his computational methodology for determination of independent-variable values.
Fig l also reveals another important aspect of Dupuy’s work, the development of reliable data on historical battles, Military judgment plays an essential role in this development of such historical data for a variety of reasons. Dupuy was essentially the only source of new secondary historical data developed from primary sources (see McQuie [1970] for further details). These primary sources are well known to be both incomplete and inconsistent, so that military judgment must be used to fill in the many gaps and reconcile observed inconsistencies. Moreover, military judgment also generates the working hypotheses for model development (e.g. identification of significant variables).
At the heart of Dupuy’s quantitative investigation of historical battles and subsequent model development is his own weapons-scoring methodology, which slowly evolved out of study efforts by the Historical Evaluation Research Organization (HERO) and its successor organizations (cf. HERO [1967] and compare with Dupuy [1979]). Early HERO [1967, pp. 7-8] work revealed that what one would today call weapons scores developed by other organizations were so poorly documented that HERO had to create its own methodology for developing the relative lethality of weapons, which eventually evolved into Dupuy’s Operational Lethality Indices (OLIs). Dupuy realized that his method was arbitrary (as indeed is its counterpart, called the operational definition, in formal scientific work), but felt that this would be ameliorated if the weapons-scoring methodology be consistently applied to historical battles. Unfortunately, this point is not clearly stated in Dupuy’s formal writings, although it was clearly (and compellingly) made by him in numerous briefings that this author heard over the years.
Figure 2.
In other words, from a system’s perspective, the functional relationships developed by Colonel Dupuy are part of his analysis system that includes this weapons-scoring methodology consistently applied (see Fig. l again). The derived functional relationships do not stand alone (unless further empirical analysis shows them to hold for any weapons-scoring methodology), but function in concert with computational procedures. Another essential part of this system is Dupuy‘s aggregation methodology, which combines numbers, environmental circumstances, and weapons scores to compute the strength (S) of a military force. A key innovation by Colonel Dupuy [1979, pp. 202- 203] was to use a nonlinear (more precisely, a piecewise-linear) model for certain elements of force strength. This innovation precluded the occurrence of military absurdities such as air firepower being fully substitutable for ground firepower, antitank weapons being fully effective when armor targets are lacking, etc‘ The final part of this computational system is Dupuy’s situational-adjustment methodology, which combines the effects of operational circumstances with force strengths to determine relative combat power, e.g. Pa/Pd.
To recapitulate, the determination of an Operational Lethality Index (OLI) for a weapon involves the combination of weapon lethality, quantified in terms of a Theoretical Lethality Index (TLI) (e.g. see Dupuy [1987, p. 84]), and troop dispersion[9] (e.g. see Dupuy [1987, pp. 84- 85]). Weapons scores (i.e. the OLIs) are then combined with numbers (own side and enemy) and combat- environment factors to yield force strength. Six[10] different categories of weapons are aggregated, with nonlinear (i.e. piecewise-linear) models being used for the following three categories of weapons: antitank, air defense, and air firepower (i.e. c1ose—air support). Operational, e.g. mobility, posture, surprise, etc. (Dupuy [1987, p. 87]), and behavioral variables (quantified as a relative combat effectiveness value (CEV)) are then applied to force strength to determine a side’s combat-power potential.
Requirement for Consistent Scoring of Weapons, Force Aggregation, and Situational Adjustment for Operational Circumstances
The salient point to be gleaned from Fig.1 and 2 is that the same (or at least consistent) weapons—scoring, aggregation, and situational—adjustment methodologies be used for both developing functional relationships and then playing them to model future combat. The corresponding computational methods function as a system (organic whole) for determining relative combat power, e.g. Pa/Pd. For the development of functional relationships from historical data, a force ratio (relative combat power of the two opposing sides, e.g. attacker’s combat power divided by that of the defender, Pa/Pd is computed (i.e. it is a derived quantity) as the independent variable, with observed combat outcome being the dependent variable. Thus, as discussed above, this force ratio depends on the methodologies for scoring weapons, aggregating force strengths, and adjusting a force’s combat power for the operational circumstances of the engagement. It is a priori not clear that different scoring, aggregation, and situational-adjustment methodologies will lead to similar derived values. If such different computational procedures were to be used, these derived values should be recomputed and the corresponding functional relationships rederived and replotted.
However, users of the Tactical Numerical Deterministic Model (TNDM) (or for that matter, its predecessor, the Quantified Judgment Model (QJM)) need not worry about this point because it was apparently meticulously observed by Colonel Dupuy in all his work. However, portions of his work have found their way into a surprisingly large number of DOD models (usually not explicitly acknowledged), but the context and range of validity of historical results have been largely ignored by others. The need for recalibration of the historical data and corresponding functional relationships has not been considered in applying Dupuy’s results for some important current DOD models.
Implications for Current DOD Models
A number of important current DOD models (namely, TACWAR and JICM discussed below) make use of some of Dupuy’s historical results without recalibrating functional relationships such as loss rates and rates of advance as a function of some force ratio (e.g. Pa/Pd). As discussed above, it is not clear that such a procedure will capture the essence of past combat experience. Moreover, in calculating losses, Dupuy first determines personnel losses (expressed as a percent loss of personnel strength, i.e., number of combatants on a side) and then calculates equipment losses as a function of this casualty rate (e.g., see Dupuy [1971, pp. 219-223], also [1990, Chapters 5 through 7][11]). These latter functional relationships are apparently not observed in the models discussed below. In fact, only Dupuy (going back to Dupuy [1979][12] takes personnel losses to depend on a force ratio and other pertinent variables, with materiel losses being taken as derivative from this casualty rate.
For example, TACWAR determines personnel losses[13] by computing a force ratio and then consulting an appropriate casualty-rate curve (referred to as empirical data), much in the same fashion as ATLAS did[14]. However, such a force ratio is computed using a linear model with weapon values determined by the so-called antipotential-potential method[15]. Unfortunately, this procedure may not be consistent with how the empirical data (i.e. the casualty-rate curves) was developed. Further research is required to demonstrate that valid casualty estimates are obtained when different weapon scoring, aggregation, and situational-adjustment methodologies are used to develop casualty-rate curves from historical data and to use them to assess losses in aggregated combat models. Furthermore, TACWAR does not use Dupuy’s model for equipment losses (see above), although it does purport, as just noted above, to use “historical data” (e.g., see Kerlin et al. [1975, p. 22]) to compute personnel losses as a function (among other things) of a force ratio (given by a linear relationship), involving close air support values in a way never used by Dupuy. Although their force-ratio determination methodology does have logical and mathematical merit, it is not the way that the historical data was developed.
Moreover, RAND (Allen [1992]) has more recently developed what is called the situational force scoring (SFS) methodology for calculating force ratios in large-scale, aggregated-force combat situations to determine loss and movement rates. Here, SFS refers essentially to a force- aggregation and situation-adjustment methodology, which has many conceptual elements in common with Dupuy‘s methodology (except, most notably, extensive testing against historical data, especially documentation of such efforts). This SFS was originally developed for RSAS[16] and is today used in JICM[17]. It also apparently uses a weapon-scoring system developed at RAND[18]. It purports (no documentation given [citation of unpublished work]) to be consistent with historical data (including the ATLAS casualty-rate curves) (Allen [1992, p.41]), but again no consideration is given to recalibration of historical results for different weapon scoring, force-aggregation, and situational-adjustment methodologies. SFS emphasizes adjusting force strengths according to operational circumstances (the “situation”) of the engagement (including surprise), with many innovative ideas (but in some major ways has little connection with previous work of others[19]). The resulting model contains many more details than historical combat data would support. It also is methodology that differs in many essential ways from that used previously by any investigator. In particular, it is doubtful that it develops force ratios in a manner consistent with Dupuy’s work.
Final Comments
Use of (sophisticated) mathematics for modeling past historical combat (and extrapolating it into the future for planning purposes) is no reason for ignoring Dupuy’s work. One would think that the current Military OR community would try to understand Dupuy’s work before trying to improve and extend it. In particular, Colonel Dupuy’s various computational procedures (including constants) must be considered as an organic whole (i.e. a system) supporting the development of functional relationships. If one ignores this computational system and simply tries to use some isolated aspect, the result may be interesting and even logically sound, but it probably lacks any scientific validity.
REFERENCES
P. Allen, “Situational Force Scoring: Accounting for Combined Arms Effects in Aggregate Combat Models,” N-3423-NA, The RAND Corporation, Santa Monica, CA, 1992.
L. B. Anderson, “A Briefing on Anti-Potential Potential (The Eigen-value Method for Computing Weapon Values), WP-2, Project 23-31, Institute for Defense Analyses, Arlington, VA, March 1974.
B. W. Bennett, et al, “RSAS 4.6 Summary,” N-3534-NA, The RAND Corporation, Santa Monica, CA, 1992.
B. W. Bennett, A. M. Bullock, D. B. Fox, C. M. Jones, J. Schrader, R. Weissler, and B. A. Wilson, “JICM 1.0 Summary,” MR-383-NA, The RAND Corporation, Santa Monica, CA, 1994.
P. K. Davis and J. A. Winnefeld, “The RAND Strategic Assessment Center: An Overview and Interim Conclusions About Utility and Development Options,” R-2945-DNA, The RAND Corporation, Santa Monica, CA, March 1983.
T.N, Dupuy, Numbers. Predictions and War: Using History to Evaluate Combat Factors and Predict the Outcome of Battles, The Bobbs-Merrill Company, Indianapolis/New York, 1979,
T.N. Dupuy, Numbers Predictions and War, Revised Edition, HERO Books, Fairfax, VA 1985.
T.N. Dupuy, Understanding War: History and Theory of Combat, Paragon House Publishers, New York, 1987.
T.N. Dupuy, Attrition: Forecasting Battle Casualties and Equipment Losses in Modem War, HERO Books, Fairfax, VA, 1990.
General Research Corporation (GRC), “A Hierarchy of Combat Analysis Models,” McLean, VA, January 1973.
Historical Evaluation and Research Organization (HERO), “Average Casualty Rates for War Games, Based on Historical Data,” 3 Volumes in 1, Dunn Loring, VA, February 1967.
E. P. Kerlin and R. H. Cole, “ATLAS: A Tactical, Logistical, and Air Simulation: Documentation and User’s Guide,” RAC-TP-338, Research Analysis Corporation, McLean, VA, April 1969 (AD 850 355).
E.P. Kerlin, L.A. Schmidt, A.J. Rolfe, M.J. Hutzler, and D,L. Moody, “The IDA Tactical Warfare Model: A Theater-Level Model of Conventional, Nuclear, and Chemical Warfare, Volume II- Detailed Description” R-21 1, Institute for Defense Analyses, Arlington, VA, October 1975 (AD B009 692L).
R. McQuie, “Military History and Mathematical Analysis,” Military Review 50, No, 5, 8-17 (1970).
S.M. Robinson, “Shadow Prices for Measures of Effectiveness, I: Linear Model,” Operations Research 41, 518-535 (1993).
J.G. Taylor, Lanchester Models of Warfare. Vols, I & II. Operations Research Society of America, Alexandria, VA, 1983. (a)
J.G. Taylor, “A Lanchester-Type Aggregated-Force Model of Conventional Ground Combat,” Naval Research Logistics Quarterly 30, 237-260 (1983). (b)
NOTES
[1] For example, see Taylor [1983a, Section 7.18], which contains a number of examples. The basic references given there may be more accessible through Robinson [I993].
[2] This term was apparently coined by L.B. Anderson [I974] (see also Kerlin et al. [1975, Chapter I, Section D.3]).
[3] The Tactical Warfare (TACWAR) model is a theater-level, joint-warfare, computer-based combat model that is currently used for decision support by the Joint Staff and essentially all CINC staffs. It was originally developed by the Institute for Defense Analyses in the mid-1970s (see Kerlin et al. [1975]), originally referred to as TACNUC, which has been continually upgraded until (and including) the present day.
[4] For example, see Kerlin and Cole [1969], GRC [1973, Fig. 6-6], or Taylor [1983b, Fig. 5] (also Taylor [1983a, Section 7.13]).
[5] The only apparent difference between Dupuy [1979] and Dupuy [1985] is the addition of an appendix (Appendix C “Modified Quantified Judgment Analysis of the Bekaa Valley Battle”) to the end of the latter (pp. 241-251). Hence, the page content is apparently the same for these two books for pp. 1-239.
[6] Technically speaking, one also has the engagement type and possibly several other descriptors (denoted in Fig. 1 as reduced list of operational circumstances) as other inputs to a historical battle.
[7] In Dupuy [1979, e.g. pp. 43-46] only environmental variables are mentioned, although basically the same formulas underlie both Dupuy [1979] and Dupuy [1987]. For simplicity, Fig. 1 and 2 follow this usage and employ the term “environmental circumstances.”
[8] In Dupuy [1979, e.g. pp. 46-47] only operational variables are mentioned, although basically the same formulas underlie both Dupuy [1979] and Dupuy [1987]. For simplicity, Fig. 1 and 2 follow this usage and employ the term “operational circumstances.”
[9] Chris Lawrence has kindly brought to my attention that since the same value for troop dispersion from an historical period (e.g. see Dupuy [1987, p. 84]) is used for both the attacker and also the defender, troop dispersion does not actually affect the determination of relative combat power PM/Pd.
[10] Eight different weapon types are considered, with three being classified as infantry weapons (e.g. see Dupuy [1979, pp, 43-44], [1981 pp. 85-86]).
[11] Chris Lawrence has kindly informed me that Dupuy‘s work on relating equipment losses to personnel losses goes back to the early 1970s and even earlier (e.g. see HERO [1966]). Moreover, Dupuy‘s [1992] book Future Wars gives some additional empirical evidence concerning the dependence of equipment losses on casualty rates.
[12] But actually going back much earlier as pointed out in the previous footnote.
[13] See Kerlin et al. [1975, Chapter I, Section D.l].
[14] See Footnote 4 above.
[15] See Kerlin et al. [1975, Chapter I, Section D.3]; see also Footnotes 1 and 2 above.
[16] The RAND Strategy Assessment System (RSAS) is a multi-theater aggregated combat model developed at RAND in the early l980s (for further details see Davis and Winnefeld [1983] and Bennett et al. [1992]). It evolved into the Joint Integrated Contingency Model (JICM), which is a post-Cold War redesign of the RSAS (starting in FY92).
[17] The Joint Integrated Contingency Model (JICM) is a game-structured computer-based combat model of major regional contingencies and higher-level conflicts, covering strategic mobility, regional conventional and nuclear warfare in multiple theaters, naval warfare, and strategic nuclear warfare (for further details, see Bennett et al. [1994]).
[18] RAND apparently replaced one weapon-scoring system by another (e.g. see Allen [1992, pp. 9, l5, and 87-89]) without making any other changes in their SFS System.
[19] For example, both Dupuy’s early HERO work (e.g. see Dupuy [1967]), reworks of these results by the Research Analysis Corporation (RAC) (e.g. see RAC [1973, Fig. 6-6]), and Dupuy’s later work (e.g. see Dupuy [1979]) considered daily fractional casualties for the attacker and also for the defender as basic casualty-outcome descriptors (see also Taylor [1983b]). However, RAND does not do this, but considers the defender’s loss rate and a casualty exchange ratio as being the basic casualty-production descriptors (Allen [1992, pp. 41-42]). The great value of using the former set of descriptors (i.e. attacker and defender fractional loss rates) is that not only is casualty assessment more straight forward (especially development of functional relationships from historical data) but also qualitative model behavior is readily deduced (see Taylor [1983b] for further details).
Dr. Peter Perla, noted defense researcher, wargame designer and expert, and author of the seminal The Art of Wargaming: A Guide for Professionals and Hobbyists, gave the keynote address at the 2017 Connections Wargaming Conference last August. The topic of his speech, which served as his valedictory address on the occasion of his retirement from government service, addressed the predictive power of wargaming. In it, Perla recalled a conversation he once had with Trevor Dupuy in the early 1990s:
Like most good stories, this one has a beginning, a middle, and an end. I have sort of jumped in at the middle. So let’s go back to the beginning.
As it happens, that beginning came during one of the very first Connections. It may even have been the first one. This thread is one of those vivid memories we all have of certain events in life. In my case, it is a short conversation I had with Trevor Dupuy.
I remember the setting well. We were in front of the entrance to the O Club at Maxwell. It was kind of dark, but I can’t recall if it was in the morning before the club opened for our next session, or the evening, before a dinner. Trevor and I were chatting and he said something about wargaming being predictive. I still recall what I said.
“Good grief, Trevor, we can’t even predict the outcome of a Super Bowl game much less that of a battle!” He seemed taken by surprise that I felt that way, and he replied, “Well, if that is true, what are we doing? What’s the point?”
I had my usual stock answers. We wargame to develop insights, to identify issues, and to raise questions. We certainly don’t wargame to predict what will happen in a battle or a war. I was pretty dogmatic in those days. Thank goodness I’m not that way any more!
The question of prediction did not go away, however.
For the rest of Perla’s speech, see here. For a wonderful summary of the entire 2017 Connections Wargaming conference, see here.
Arnold Schwarzenegger and friend. [Image Credit Jordan Strauss/Invision/AP/File]
Humans are a competitive lot. With machines making so much rapid progress (see Moore’s Law), the singularity approaches—see the discussion between Michio Kaku and Ray Kurzweil, two prominent futurologists. This is the “hypothesis that the invention of artificial super intelligence (ASI) will abruptly trigger runaway technological growth, resulting in unfathomable changes to human civilization.” (Wikipedia). This was also referred to as general artificial intelligence (GAI) by The Economist, and previously discussed in this blog.
We humans also exhibit a tendency to anthropomorphize, or to endow any observed object with human qualities. The image above illustrates Arnold Schwarzenegger sizing up his robotic doppelgänger. This is further evidenced by statements made about the ability of military networks to spontaneously become self-aware:
The idea behind the Terminator films – specifically, that a Skynet-style military network becomes self-aware, sees humans as the enemy, and attacks – isn’t too far-fetched, one of the nation’s top military officers said this week. Nor is that kind of autonomy the stuff of the distant future. ‘We’re a decade or so away from that capability,’ said Gen. Paul Selva, vice chairman of the Joint Chiefs of Staff.
This exhibits a fundamental fear, and I believe a misconception, about the capabilities of these technologies. This is exemplified by Jay Tuck’s TED talk, “Artificial Intelligence: it will kill us.” His examples of AI in use today include airline and hotel revenue management, aircraft autopilot, and medical imaging. He also holds up the MQ-9 Reaper’s Argus (aka Gorgon Stare) imaging systems, as well as the X-47B Pegasus, previously discussed, as an example of modern AI, and the pinnacle in capability. Among several claims, he states that the X-47B has an optical stealth capability, which is inaccurate:
[X-47B], a descendant of an earlier killer drone with its roots in the late 1990s, is possibly the least stealthy of the competitors, owing to Northrop’s decision to build the drone big, thick and tough. Those qualities help it survive forceful carrier landings, but also make it a big target for enemy radars. Navy Capt. Jamie Engdahl, manager of the drone test program, described it as ‘low-observable relevant,’ a careful choice of words copping to the X-47B’s relative lack of stealth. (Emphasis added).
Such questions limit the veracity of these claims. I believe that this is little more than modern fear mongering, playing on ignorance. But, Mr. Tuck is not alone. From the forefront of technology, Elon Musk is often held up as an example of commercial success in the field of AI, and he recently addressed the national governors association meeting on this topic, specifically in the need for regulation in the commercial sphere.
On the artificial intelligence [AI] front, I have exposure to the most cutting edge AI, and I think people should be really concerned about it. … AI is a rare case, I think we should be proactive in terms of regulation, rather that reactive about it. Because by the time we are reactive about it, its too late. … AI is a fundamental risk to human civilization, in a way that car crashes, airplane crashes, faulty drugs or bad food were not. … In space, we get regulated by the FAA. But you know, if you ask the average person, ‘Do you want to get rid of the FAA? Do you want to take a chance on manufacturers not cutting corners on aircraft because profits were down that quarter? Hell no, that sounds terrible.’ Because robots will be able to do everything better than us, and I mean all of us. … We have companies that are racing to build AI, they have to race otherwise they are going to be made uncompetitive. … When the regulators are convinced it is safe they we can go, but otherwise, slow down. [Emphasis added]
Mr. Musk also hinted at American exceptionalism: “America is the distillation of the human spirit of exploration.” Indeed, the link between military technology and commercial applications is an ongoing virtuous cycle. But, the kind of regulation that exists in the commercial sphere from within the national, subnational, and local governments of humankind do not apply so easily in the field of warfare, where no single authority exists. Any agreements to limit technology are a consensus-based agreement, such as a treaty.
The husky was mistakenly classified as wolf, because the classifier learned to use snow as feature. [Machine Master blog]
In a recent TEDx talk, Peter Haas describes his work in AI, and some of challenges that exist within the state of the art of this technology. As illustrated above, when asked to distinguish between a wolf and a dog, the machine classified the Husky in the above photo as a wolf. The humans developing the AI system did not know why this happened, so they asked the AI system to show the regions of the image that were used to make this decision, and the result is depicted on the right side of the image. The fact that this dog was photographed with snow in the background is a form of bias – are fact that snow exists in a photo does not yield any conclusive proof that any particular animal is a dog or a wolf.
Right now there are people – doctors, judges, accountants – who are getting information from an AI system and treating it like it was information from a trusted colleague. It is this trust that bothers me. Not because of how often AI gets it wrong; AI researchers pride themselves on the accuracy of results. It is how badly it gets it wrong when it makes a mistake that has me worried. These systems do not fail gracefully.
AI systems clearly have drawbacks, but they also have significant advantages, such as in the curation of shared model of the battlefield.
In a paper for the Royal Institute of International Affairs in London, Mary Cummings of Duke University says that an autonomous system perceives the world through its sensors and reconstructs it to give its computer ‘brain’ a model of the world which it can use to make decisions. The key to effective autonomous systems is ‘the fidelity of the world model and the timeliness of its updates.‘ [Emphasis added]
Perhaps AI systems might best be employed in the cyber domain, where their advantages are naturally “at home?” Mr. Haas noted that machines at the current time have a tough time doing simple tasks, like opening a door. As was covered in this blog, former Deputy Defense Secretary Robert Work noted this same problem, and thus called for man-machine teaming as one of the key areas of pursuit within the Third Offset Strategy.
Just as the previous blog post illustrates, “the quality of military men is what wins wars and preserves nations.” Let’s remember Paul Van Ripper’s performance in Millennium Challenge 2002:
Red, commanded by retired Marine Corps Lieutenant General Paul K. Van Riper, adopted an asymmetric strategy, in particular, using old methods to evade Blue’s sophisticated electronic surveillance network. Van Riper used motorcycle messengers to transmit orders to front-line troops and World-War-II-style light signals to launch airplanes without radio communications. Red received an ultimatum from Blue, essentially a surrender document, demanding a response within 24 hours. Thus warned of Blue’s approach, Red used a fleet of small boats to determine the position of Blue’s fleet by the second day of the exercise. In a preemptive strike, Red launched a massive salvo of cruise missiles that overwhelmed the Blue forces’ electronic sensors and destroyed sixteen warships.
We should learn lessons on the over reliance on technology. AI systems are incredibly fickle, but which offer incredible capabilities. We should question and inspect results by such systems. They do not exhibit emotions, they are not self-aware, they do not spontaneously ask questions unless specifically programmed to do so. We should recognize their significant limitations and use them in conjunction with humans who will retain command decisions for the foreseeable future.
.PNG)
