Discover millions of ebooks, audiobooks, and so much more with a free trial

Only $11.99/month after trial. Cancel anytime.

Differential Games: A Mathematical Theory with Applications to Warfare and Pursuit, Control and Optimization
Differential Games: A Mathematical Theory with Applications to Warfare and Pursuit, Control and Optimization
Differential Games: A Mathematical Theory with Applications to Warfare and Pursuit, Control and Optimization
Ebook748 pages5 hours

Differential Games: A Mathematical Theory with Applications to Warfare and Pursuit, Control and Optimization

Rating: 4 out of 5 stars

4/5

()

Read preview

About this ebook

One of the definitive works in game theory, this fascinating volume offers an original look at methods of obtaining solutions for conflict situations. Combining the principles of game theory, the calculus of variations, and control theory, the author considers and solves an amazing array of problems: military, pursuit and evasion, games of firing and maneuver, athletic contests, and many other problems of conflict.
Beginning with general definitions and the basic mathematics behind differential game theory, the author proceeds to examinations of increasingly specific techniques and applications: dispersal, universal, and equivocal surfaces; the role of game theory in warfare; development of an effective theory despite incomplete information; and more. All problems and solutions receive clearly worded, illuminating discussions, including detailed examples and numerous formal calculations.
The product of fifteen years of research by a highly experienced mathematician and engineer, this volume will acquaint students of game theory with practical solutions to an extraordinary range of intriguing problems.
LanguageEnglish
Release dateApr 26, 2012
ISBN9780486135984
Differential Games: A Mathematical Theory with Applications to Warfare and Pursuit, Control and Optimization

Related to Differential Games

Titles in the series (100)

View More

Related ebooks

Mathematics For You

View More

Related articles

Reviews for Differential Games

Rating: 4 out of 5 stars
4/5

2 ratings0 reviews

What did you think?

Tap to rate

Review must be at least 10 words

    Book preview

    Differential Games - Rufus Isaacs

    CHAPTER 1

    An Introduction

    An informal introduction to the subject and some of its applications, this chapter precedes the formal mathematical development.²

    1.1. THE THEORY OF GAMES

    It is only a few years since J. von Neumann and O. Morgenstern innovated the theory of games.² Their work gave us such essential items as the concept of a strategy, the value of a game, and, probably most important of all, a workable and sound delineation of an optimal strategy which might be either pure or mixed.

    The reader of this book should be conversant with these terms as they relate to zero-sum, two-player games, but further technical knowledge of game theory will be needed but rarely. These researches will be essentially self-contained; although they could not have existed without the above pioneering work, its ideas for us will be standards and models rather than tools.

    Game theory seemed to leap at once to the status of an accepted classic. The subject was innundated by a flood of papers and a wave of books. Optimism ran high at certain periods as to the revolution pending in certain domains of its applications, predominantly warfare and economics.

    In the former, which we are far better equipped to discuss, very little happened. What are the reasons for the failure of such high hopes?

    It seems there are two. One is the increased difficulty of the problems when—and such is the essence of game theory—there are two opponents with conflicting aims and each is to make the best possible decisions understanding and taking into account that his antagonist is doing the same. Such a situation is entirely different from what we might call the corresponding class of problems in classical analysis. As we shall discuss further in Chapter 3, it is possible to cast many of these problems in the form of one-player games. The added difficulty with two interacting players can be enormous.

    The second reason is the lack of methods for obtaining answers. Of the deluge of material written on game theory, most concerns general theorems and results, often of the highest calibre mathematics but very little of usable techniques for obtaining practical answers. Such, as the previous paragraph states, is not always easy, but even so it is seldom the goal. Mathematics today tends to favor more abstract and general ends.

    These ideas are discussed further in Chapter 11 in connection with military applications of our subject. Except for certain technical sections, most of this chapter can be read earlier than in context, as a sort of supplementary introduction.

    Almost all elementary expositions of game theory begin with the game matrix and its attendant concepts. This is an excellent device for formulating and proving the basic theorems without which the subject could not exist, but—and such does not appear adequately well realized—aside from certain very simple instances, the matrix is not a satisfactory tool for solving games. It is a formulation, not an end, and for such purposes as obtaining answers, a very unsatisfactory one.³

    Unless the game is extremely simple—say each player’s decision consists of one choice from either a small number of discrete alternatives or from a continuum with an elementary mathematical structure—the matrix will be astronomically large.

    If a method for solution is to exist in the sense of mathematical analysis, the game must have a certain inner logical pattern that will yield to such analysis. We have always found this idea difficult to make clear. It is amplified in Chapter 3, which the reader can turn to now if he likes.

    A definition of differential games should hinge on this concept. The name itself suggests that we are going to approach problems in game theory through such devices of classical analysis as differential equations. So, for the most part, we are. But we prefer to think of our subject as that which treats games in which both players are confronted with lengthy sequences—be they continuous or discrete—of decisions which are knit together logically so that a perceptible and calculable pattern prevails throughout.

    1.2. THE STATE AND CONTROL VARIABLES

    Certain types of battles, airplanes dog-fighting, football, a torpedo pursuing a ship, a missile intercepting an aircraft, a gunner guarding a target against an invader, are typical models for differential games. If one of the two players is suppressed, the theory becomes one of maximization. It relates to the calculus of variations and subsumes the major portion of control theory.

    Always the players make their decisions through choosing the values of certain control variables. These, in turn, govern the values of certain other quantities called state variables. The latter are such that if their values are known at any instant, the current state of the game is fully determined thereby. To be clear we will state this concept in three equivalent ways.

    State variables enjoy the properties:

    Their values must be known at the outset in order to determine the outcome of a game.

    They are exactly the values that are relevant at each instant to a player making decisions as to how to play.

    If a player had to be replaced by a substitute during a partie, the information the latter would require to resume play would be just the current values of the state variables.

    As the play of a game progresses, the state variables change. If there are n of them so that they may be designated by x1, . . . , xn, we may regard these quantities as the coordinates of a point in Euclidean n.

    The motion of x (= {x1, . . . , xn}) is at all times under the partial control of the two players, a control which is manifested through their choice of the control variables.

    The current values of the state variables are always known to both players; thus we treat games with full information. Such perhaps is the greatest limitation on the present form of the theory especially in the domain of military strategy. The extension to cases of partial information appears to be the most vital area for future research. Chapter 12, which may also be read out of consecutive order, is devoted to promising ideas of such developments.

    . It terminates when certain conditions are fulfilled, and it is always possible to arrange matters so that these conditions consist of x’s being on a certain surface or (n .

    When the partie is thus completed, there will be known the numerical value of a quantity called the payoff. It will be the objective of one player to maximize and of his opponent to minimize it. As in standard game theory the best value of the payoff, its minimax, will be termed the Value (we shall always capitalize it) of the game. It will be the payoff when both players act optimally. Should one fail to do so, there will be a way for his opponent to attain a payoff more favorable to him than the Value.

    But let us view these concepts now in terms of one general class of applications.

    1.3. BATTLE GAMES

    The state variables should be such quantities (for both sides) that are indicative of the current state of affairs as is relevant to the model selected, which must necessarily be a simplified version of reality. These will include such items as numbers of men, aircraft, tanks, ships and other munitions, possibly subdivided by allocation to different sectors or by classification into different types, the number of miles advanced by a front, and so on.

    Let Army 1 be minimizing and have jurisdiction over the control variables, φ1, . . . , φλ; similarly, Army 2 maximizes and has the control variables, ψ1, . . . , ψk. These roles are, of course, matters of arrangement. Suppose, for example, the payoff was to be the excess of men (or munitions or any other vital material) at the end of, say, a stipulated time T. If xi is the number of such then possessed by Army i, we take for the payoff

    x2 − x1.

    The mechanism by which a partie progresses can perhaps be gleaned from some partial instances. The xi will be typical state variables.

    Suppose x1 represents the number of men at a certain sector for Army 1. Enemy air raids may deplete this number. Suppose x3 is the number of Army 2’s aircraft available for this purpose and he decides to send over at a certain time a fraction ψ1 of them. We must now decide from experience or otherwise how the expected number of casualties depends on the number ψ1x3 of present enemy aircraft. Let us suppose it is directly proportional with a constant c picked empirically.

    To be able to exploit the powerful tools of mathematical analysis, we will suppose continuous processes rather than discrete steps. Such gives a long-term, but rounded-out picture.

    In the case just given let us also suppose Army 1 furnishes replacements to the post in question at a fixed rate r; then an equation begins to emerge:

    (1.3.1)

    The dots on the right signify various other terms, such as more negative ones arising from other means of attack by Army 2 or the deployment of the men for other purposes by Army 1 and positive terms such as may arise from defensive measures of 1 against 2’s attacks. If the game is at all symmetrical, there, of course, will be a similar equation with the roles of the players reversed.

    Again let x4 be Army 1’s stock of a certain munition, which is in short supply, so that economy in its use is critical. Let b be the maximal firing rate of this munition. Let φ2 (0 ≤ φ2 ≤ 1) be the fraction of b that 1 decides to use at any time. We will have

    (1.3.2)

    , we shall demand that x4 ≥ 0. Then (1.3.2) implies a constraint on the munition and gives the player scope for distributing its consumption within this constraint.

    In both (1.3.1) and (1.3.2) the left sides are to be interpreted as forward time derivatives. Equations of this type are typical of the underlying means of progress of a differential game. In general, they are called the kinematic equations (KE) and have the form

    (1.3.3)

    Thus the forward rates of changes of the state variables are given functions of the state and control variables (the latter are the φ and ψ) of both players.

    1.4. GAMES WITH MOVING CRAFT

    To display the kinematic equations, state and control variables, and the distinction between the latter two, we shall pick, as an instance of a moving craft, an automobile. Our motive is simply that here is a vehicle whose behavior is known to all. The ideas hold, with at most small changes, for almost any craft or vehicle; we can apply them to tanks or ships—outboards to cruisers. Aircraft, of course, move in three-dimensional space, but the principles are the same.

    Three state variables at first appear to be needed to describe the auto’s geometrical position, namely x1 and x2, the Cartesian coordinates of some fixed point on the car, and x3, the inclination or direction in which the car is pointing. That is, x3 is the angle between the main axis of the car and the x1-axis as drawn on the unbounded, empty parking lot which we will take for our theatre of operations. But if the auto is to figure in a differential game, we need more. Let us assume that the car is navigated by an accelerometer and steering wheel.⁵ The former we will suppose controls the tangential acceleration. This quantity, being under the control of a player, will be a control variable and might be denoted by φ1. But, as it proves convenient to have simple uniform bounds on the control variables, we take the forward acceleration as 1. Here A is the maximal possible acceleration, and the control variable φ1 can now be constrained: 0 ≤ φ1 ≤ 1. It is thus the fraction of the full acceleration and under the driver’s control. The speed x4 is not directly under the control of the driver, but it is a quantity both players of a game involving the car would have to take into account as much as x1, x2, x3. Hence it, too, will be considered a state variable.

    The position of the steering wheel determines the curvature of the car’s path. But to say that the driver can vary it from instant to instant is not realistic. Rather we will take the curvature of the car’s path as another state variable x6 (physically, it is visible from the inclination of the front wheels) and its fractional rate of change as another control variable φ2. That is, if W is the maximal possible rate, the rate selected by the driver will be 2 with -1 ≤ φ2 ≤ 1.

    Under these assumptions, the motion of the automobile will be governed by the following kinematic equations:

    (1)

    (2)

    (3)

    (4)

    (5)

    Here (1), (2) is simply the decomposition of the car’s velocity into its axiswise components; (3) states that the rate of change of direction is the speed times the curvature. As to (4), the rate of change of speed is the acceleration and (5) has been discussed.

    To recapitulate, x1, . . . , x5 describe the aspects of our vehicle which would be relevant were it a participant in, say, a pursuit game and are called state variables. The driver controls φ1 (the accelerometer) and φ2 (the fractional speed at which he turns the steering wheel). They are the control variables, and they alone are at each instant under the control of the player. They are not measurable by the opponent; the state variables are.

    The reader will at once perceive the deficiencies in our model. The most glaring is that there is no check on the speed. This could be remedied by putting a bound on x4, but a more realistic diagnosis would lay the blame on (4). First, it is probably oversimplifying automobile dynamics to state that the force exerted by the motor is proportional to the amount we depress the accelerator pedal; secondly, and more important, this force is proportional to the acceleration of the car only if we neglect friction. If we assume, for the sake of simplicity, that friction is negatively proportional to the speed, a better version of (4) would be

    4 = F(1) — Kx4.

    Here 1 (0 ≤ φ1 ≤ 1) is the amount the accelerator pedal is depressed,⁶ F the resulting force (per unit mass of the car) exerted by the engine, and K a friction coefficient. The speed now will be bounded by F(A)/K.

    Another essential correction consists of bounding the curvature x5. (Anyone who has tried to turn his car in a narrow street needs be told no more.)

    Thus the equations of motions may be complicated for a closer simulation of reality or simplified for easier mathematics. In most of the sample problems to be given in later chapters we will modify in the latter direction, for at times the mathematics will be heavy enough without refinements of detail which add little of principle.

    Let us take a second example. Here a point will move in the plane with no other restriction⁸ save that its speed w is constant. The controlling player picks the direction of travel and may change it abruptly at any instant. There is a single control variable, φ, the direction. The KE are

    (1.4.1)

    Here we have but two state variables, x1 and x2. This situation will be spoken of as simple motion.

    1.5. PURSUIT GAMES

    The various guises pursuit games may assume in warfare are legion-torpedo and ship, ship and submarine, missile and bomber, tank and jeep.

    To obtain a very general picture we will denote the pursuer by P and the evader by E. The craft they represent can be steered by either a human pilot or an automatic mechanism. In more complex versions we may have more craft, such as several small fighter planes opposing an enemy fleet of bombers or—to leap to a new arena—several tacklers against a ball carrier with interference runners in a football game. In the general sense, P and E represent the players—the opposed controlling minds of the conflict. But when each governs but one craft, these symbols will usually designate the vehicles themselves. Thus P, say, will be some fixed point on the pursuing craft so that the coordinates of P specify the craft’s geometrical location (but not its position, a term which will also entail heading, speed, and all such quantities which are state variables).

    A pursuit game usually terminates when capture occurs, which means that the distance PE becomes less than a certain prescribed positive quantity l.

    To clarify our ideas let us settle on some definite typical instance. For E we take an invading bomber, either a plane or guided missile, and for P a defending interceptor, also either a plane or missile. First we ask the question: How best should P pursue E? That is, if at each instant of time P knows his own and E’s position, how should P at this instant regulate the various control or steering variables at his disposal? By position we mean not only the geometric location of P or E but also all other relevant magnitudes, such as flight direction, orientation, and velocity, in short, the relevant state variables.

    Secondly, we must decide what is meant by best. In terms of game theory, we must select a payoff. The most obvious criterion is whether capture can be achieved at all. In such a case, where we are interested only in two (or any finite number) outcomes, we shall speak of the problem as a game of kind (in contrast to games of degree, which have a continuum of outcomes). But P may be an interceptor with a limited fuel supply. Then a more realistic criterion would be based on whether capture can be achieved prior to the elapse of a certain stipulated time. If E is a bomber whose objective is to reach a certain target, the point of interest may be whether capture can be attained before E reaches his destination. If P is to use a gun, rocket, or some such weapon, capture will consist of bringing E within range. If there is but a probability of a hit less than certainty, P may wish to keep E within range for a stipulated period.

    All the above criteria are discrete, or rather of a yes-or-no variety, and we catalogue them as games of kind. But there are also cases where there is some variable quantity which the opponents conflictingly seek to maximize or minimize. This quantity is the payoff and the game becomes one of degree.

    It is often possible to find such a continuous payoff in such a way that the aforementioned discrete criteria are automatically subsumed within it. For example, suppose we are interested only in whether capture can be achieved or not. We can pick as payoff the time of capture, with P’s objective being to make this quantity as small as possible and E’s, as great. No capture at all corresponds to infinite time. Then if P acts in accordance with this dictate, he will certainly achieve his primary objective of just capture whenever possible. And more. He will do it as rapidly as possible. Again, suppose the objective was initially capture before a stipulated time T. By minimizing the time of capture, P will surely succeed if it is possible for him to do so. We need but look at the minimal value of the capture time that P can attain and see whether or not it exceeds T.

    This idea is fairly general. If the original desideratum was whether or not E could attain a certain proximity to a certain target, we can make the payoff the distance from the target when capture takes place. By having P strive to maximize this quantity, we are assured that he will not only attain his objective of protecting the target when possible but also the biggest margin of safety, or smallest deficiency if he cannot frustrate E.

    We answer our question as to what is meant by best in all cases by deciding on a numerically valued payoff. For games of kind we could do this a bit artificially by assigning two (or more) numerical values to be the payoff for the two (or more) outcomes. Best for P means making this payoff as small⁸ as possible.

    Supposing that a payoff is elected, how shall P minimize it? If he is pursuing a missile E, how should he act? Should he, for example, using data culled from measurements of E’s position, endeavor to extrapolate E’s future course and maneuver so as to head him off?

    A brief reflection shows that such questions are meaningless. Their answer depends on how E is going to behave. If he adopts the naive policy of traveling in a straight line at constant speed, then of course P should head him off, and it is a simple matter to compute the best way of his doing so. But should P always act so, E, if he is astute, can easily frustrate P by feinting false directions and decoying P into false predictions. No plan of pursuit for P will be optimal under all types of opposition.

    Figure 1.5.1

    It is here we see that the theory of games enters and permeates the subject. We cannot talk of optimal pursuit without also speaking of optimal evasion. The behavior possibilities for both opponents must be considered together before we can evolve a technique for analyzing the situation. Such is precisely what is done by the present theory of differential games.

    Optimal evasion then enjoys a status ranking with optimal pursuit. All the remarks made a few paragraphs back about P’s+ objectives for pursuit have their counterparts about E’s objectives for evasion. For example, we might have (and indeed should have) spoken of E’s methods for avoiding capture or at least forestalling it until the elapse of T. We might have spoken, in case distance from the target at capture was the payoff, of how E is to maximize it. In actual warfare, of course, both sides will have to consider both classes of questions. It was only to ease our exposition that we wrote the earlier discussion of objectives and payoffs from P’s point of view.

    The reader may enjoy the following simple pursuit game, whose solution will appear later in this chapter. The answer can be ascertained by elementary geometrical reasoning; the method is not typical of the more general ones we shall subsequently develop. Simple as is the reasoning required—for the problem is deliberately so contrived—very few persons hit on the correct answer.

    In Figure 1.5.1, C is a target area which P is guarding from the enemy attacker E. Both P and E travel with simple motion with the same speed and start from the positions shown. For simplicity, here we shall take capture to mean the coincidence of P and E. The payoff is to be the distance from the point of capture, if any, to C, which P is to maximize and E to minimize. If E can reach C without being captured, he regards this outcome as best of all. How should each craft travel?

    We can think of E as carrying a powerful weapon, say a nuclear warhead, and, if he cannot reach the target, he at least wants detonation to be as close to it as possible. Accordingly, the interceptor P endeavors to meet E at a point as far as possible from C.

    We present a second example, very far from simple. It is intended as a pursuit game with one of the combatants suffering a proscription on his curvature. Such, of course, is a basic type of kinematic constraint; later we shall treat the case where both craft are so restricted, but such complicates the problem without sufficient compensation here in the form of new principles.

    Although this problem typifies certain general pursuit games, the macabre title (see the ensuing example) is perhaps an aid to succinct visualization. We can think of an automobile on an infinite unobstructed parking lot attempting to run down a pedestrian. The pith of the idea is, of course, a pursuer with higher speed but with the disadvantage of poor mobility.

    This simple pursuit game embodies such a rich assortment of typical phenomena that it has often served as a guidepost in the construction of the present theory. Because it can do the same for the reader’s comprehension, we introduce it here. Repeatedly, in the later sections of this book, we shall use it as an example and guinea pig.

    It is simple enough for the action to be readily envisioned, yet a partie may entail a diverse sequence of stages, some of them far from obvious. There is an opulence of singular surfaces,⁹ many displayed with pristine typicality. Yet it is possible to interpret many facets geometrically and partially solve the game without analysis. A comparison clarifies and verifies our ideas; the geometric version, as well as the completion of the full solution, will appear in Chapter 10.

    Example 1.5.1. The homicidal chauffeur game. The action takes place in the plane. The pursuer P moves at a fixed speed w1 but with his radius of curvature bounded by a given quantity R. He steers by selecting the value of this curvature at each instant. Such is a rough simulation of the navigation scheme of an automobile, boat, or aircraft; control is through the position of the steering wheel, idealized in that its setting can be instantaneously changed.

    The evader E moves with simple motion. That is, his speed w2 is fixed and he steers by, at each instant, choosing his direction of travel. Abrupt changes in this choice are allowed; his path does not everywhere have to possess a tangent.

    Capture occurs when the distance PE ≤ a given quantity /, the capture radius. The pursuer is to be the faster: w1 > w2.

    We are interested in two problems.

    1. The game of kind. When can P catch E? Clearly if R is large enough, l small, and w1 not greatly exceeding w2, E can always escape. We can envisage him doing so by merely sidestepping whenever capture appears imminent. For the limitation on P’s curvature can prohibit a sharp enough turn; he can but streak past E and, after steering to a position for a new try, be defeated again by the same maneuver.

    The problem is to find the precise conditions, values of R, l, w1/w2, which demarcate this possibility. It will be solved in Example 9.1 and Chapter 10.

    2. The game of degree with time of capture as the payoff. Now we suppose that P can always achieve capture and adopt as payoff the time it takes. In terms of our gruesome parable, we may envisage the pedestrian as hoping for the arrival of rescuers and so, if he cannot deter capture, he at least endeavors to defer it. Similarly, P strives for as quick a conclusion as possible.

    When initially E is more or less in front of P, optimal play is more or less obvious. At (a) of Figure 1.5.2, P is initially at P, his velocity being upward ; E is at E, in front of P and, say, a little to his right. Also drawn is part of P’s right circle of maximal curvature, tangent at P to the velocity vector. Under the dictates of his optimal strategy, P should begin by traversing this circle—the sharpest possible right turn—to P1 where his velocity points at E. From there he traverses this tangent as shown. Similarly, E travels this tangent, and this simple pursuit along it continues until capture occurs, say, at C.

    But now let E start from a position somewhat rear of E as shown at (b). If P plays as above, it may well be that E has time to get well within the curvature circle before P’s arrival. Clearly P’s maneuver was futile.

    To capture he must act less directly, as at (c). First he goes away from E and, when far enough, wheels about to commence a more direct pursuit (see the path sketched). On the other hand, E, realizing that time is the payoff, strives to delay P’s achieving an adequate distance PE. To this end, he begins his flight by following P, say along EE1 in the figure. At some point E1, he turns tail and flees, the partie concluding with a direct chase as at (a).

    Figure 1.5.2

    This latter indirect pursuit will be referred to as a swerve.¹⁰ It constitutes the more mathematically interesting aspect of the game of degree. Quantitatively, what is the solution and how is it found? Exactly what are the optimal paths of P and E? We must await Chapter 10 for an answer.

    1.6. GAMES OF KIND AND GAMES OF DEGREE

    We have mentioned how, in pursuit games, it is at times wiser to imbed a game of kind in one of degree. The idea is, of course, general.

    The objective of a battle may be of some concrete, yes-or-no, type. We have a game of kind. For the payoff we select two numbers corresponding to whether or not the objective is attained, the smaller number pertaining to the minimizing side’s desideratum. As before, we may advantageously turn the game of kind into one of degree. For example, a side may endeavor to attain the objective with a minimum expenditure of some item. The payoff will then generally be of the form of an integral of some quantity extending over time from the beginning to the end of the play.

    In a war of extermination, the partie ends when one side is annihilated. For a payoff we might take the number of survivors of the victorious side.

    1.7. STRATEGIES

    In conventional game theory, a strategy for a player consists of a set of decisions which tell what move he will make for every possible position that may arise in the course of a partie. If each player selects a strategy, the outcome of the game is completely determined thereby.

    The natural analogue in differential games is the selection of the control variables as functions of the state variables. Thus things are as they should be: for each position that may arise, that is, set of values of the state variables, each player has decided on his course of action, that is, he has chosen values for his control variables. It is easy to see that here, too, the outcome is determined. Let P select the φj and E select the ψj as functions of the xi.¹¹ If these functions are reasonably simple and they are substituted in the kinematic equations (1.3.3), the right side becomes a function of the xj. Then the KE become a set of simultaneous ordinary differential equations. They are to be integrated, employing as initial conditions the particular values of xj at which the play starts. The solutions will give the xj as functions of the time t and will describe the action that ensues from the chosen strategies.

    The payoff can now be computed. We find ourselves in a game theoretic terrain. The objectives of the players are to pick the φj(x) and ψj(x) so as to minimize and maximize the payoff, respectively.

    Although this heuristic description of the mechanism of our theory consorts well with practice, in the next chapter we shall give a sharper definition of a strategy. The principle will remain the same, but certain difficulties, which the mathematician will at once perceive, will be alleviated.

    In pursuit problems, a strategy admits a direct, mechanistic interpretation. An automatic guidance system must comprise, first, a detection device to obtain the relevant information about the opponent’s current status and, secondly, a decision device which regulates the steering controls in accordance with this information. An essential question is how best to design the decision mechanism. The answer is immediate and striking. A design scheme for the decision device is exactly a choice of a strategy in the sense just discussed. For what the detection device does is to ascertain the values of the state variables; what the decision device does is to select values of the control variables dependent on them. Thus a design scheme simply prescribes that the control variable will be certain functions of those of state. But this is precisely what we termed a strategy.

    1.8. DOGFIGHTS, FIRING GAMES, PROGRAMMING, AND ATHLETICS

    The scope of differential games as covered in this book can probably best be glimpsed by skimming through the volume, but we add some desultory comments here.

    A dogfight is to entail aircraft with guns that can be aimed only by pointing the entire plane. The criterion would depend on which craft gets his antagonist in range first. With the motion governed by a suitable set of kinematic equations, such would be a game of kind. An alternative is a continuous payoff, the hit probability (or differences thereof for the two players), expressed in terms of some sort of time integral extended over the periods when one player has the other in range.

    Another class of military games has a fixed battery of some type firing at a moving target. The strategies of the former might involve the allocation of a limited amount of ammunition; those of the latter, kinematic maneuvering. The most likely payoff would be the hit probability (see Section A5 of the Appendix).

    Our methods are applicable to certain minimizing problems.¹² All we need do is consider one player inactive, that is, with no control variables. We are thus provided with a tool for a certain class of problems where the goal is an optimal program of operations. Such occur in the field known as control theory. Indeed, the formalization of this science employs equations which become identical to our KE when, say, the ψi are suppressed. An instance is furnished by Section A4 of the Appendix. An operations research type optimizing problem is illustrated by Example 5.4. How our methods may obtain optimal trajectories of guided missiles is shown in Appendix A3.

    There are times in such cases when we are led to results analytically equivalent to those of the calculus of variations, even though our approach seems quite distinct. We shall note the comparison at several points in future chapters.

    On the other hand, we have been able to take some of the hoary classical problems of the variational calculus and, by adding an opponent, make two-player games of them.

    We can fit our theory nicely to certain athletic games. Football, for example, is a natural. A single ball carrier opposed by a single tackler is nothing other than a simple pursuit game with distance from the goal to the point of capture as payoff. The situation is still in our scope if there are several tacklers and also interference runners.

    1.9. TWO EXAMPLES

    Both are rather contrived pursuit games, but they are

    Enjoying the preview?
    Page 1 of 1