Você está na página 1de 12

AIAA Guidance, Navigation, and Control Conference 08 - 11 August 2011, Portland, Oregon

AIAA 2011-6712

Near-Optimal Trajectory Shaping of Guided Projectiles with Constrained Energy Consumption

N. Lchevin1 and C.A. Rabbath2 Defence R&D Canada Valcartier, Quebec City, Canada G3J 1X5

An iterative search based on cross-entropy minimization is proposed to generate a sequence of lateral acceleration signals that command a guided projectile to a prescribed target set. The proposed approach is based on iterative simulations, thus enabling the projectile to satisfy, when possible, tight terminal constraints while taking into account nonlinear flight dynamics. To do so, a trajectory shaper is proposed. The trajectory shaper is composed of a discrete-valued lateral acceleration command in series with a smoothing filter. Non-gradient-based iterative searches are utilized to quickly determine the sequence of lateral accelerations that fulfill the control objectives. Simulations show that the crossentropy method achieves near optimality in a satisfactory period of time. Furthermore, it is shown that energy expenditure can be limited while meeting the target set by simply including a carefully designed smoothing filter rather than trying to minimize a more complex objective function.

I. Introduction
odern precision weapons aim at reducing collateral damage both in terms of infrastructure and loss of life, and at providing timely and efficient debilitation of the target by the use of a small number of rounds that are quickly fired and accurately guided. To be effective in an operational theatre, calculation of the optimal trajectory needed to guide the precision weapon must be promptly achieved. In the context of precision guided munitions, compliance with time-critical operating conditions, and with time-varying as well as partially known disturbances such as atmospheric turbulence, calls for a trajectory shaping solution that integrates up-to-date information prior to firing the rounds. The challenges that arise in deriving a fast trajectory shaping solution increase with the desire to satisfy stringent control constraints, such as actuator saturation, and the presence of nonlinearities inherent with such dynamic system. The trajectory shaping solution proposed in this paper consists of calculating, prior to launch, a lateral acceleration-time pair that the controlled projectile can robustly track by means of an appropriate guidance law and autopilot. Alternatively, a position-time (waypoint-time) pair can be obtained prior to firing the weapon. Trajectory shaping can be interpreted as a constrained feedforward controller that drives the projectile to a prescribed target set under nominal operating conditions. The target set is defined by a number of constraints that include the projectiles range, terminal speed, and terminal flight-path angle. Such constraints may be needed to reduce collateral damage. For example, a projectile may be required to hit a target in an orthogonal orientation with respect to the ground. The guidance law is assumed to provide, during the course of the flight, some robustness when operating conditions differ from the nominal ones. Several approaches can tackle this constrained control design problem. Backward reachable set theory1 and the related invariant set theory have been used to solve the trajectory shaping problem. Indeed by deriving an invariant set from the prescribed target set, every trajectory that starts in the invariant set, if it exists, is guaranteed to be captured by the target set1. Application of level-set theory1,2,3 and viability theory4,5 allows derivation of an invariant set, although at the expense of fast computation since it typically requires solving time-dependent Hamilton-JacobiIsaacs partial differential equations. Differential flatness is another concept that can be used to derive commands that enable a dynamic system to follow a prescribed trajectory that aims to reach a target set6. To do so, a flat output, possibly a nonlinear function of the projectiles position, speed, and flight-path angle, is found so that the lateral acceleration command can be expressed as a function of the flat output and its time derivatives. This approach may not be appropriate when the trajectory shaping problem includes consideration of control saturation since it leads to
1 2

Defence Scientist, nicolas.lechevin@drdc-rddc.gc.ca. Defence Scientist, camille-alain.rabbath @drdc-rddc.gc.ca. 1 American Institute of Aeronautics and Astronautics

Copyright 2011 by the authors. Published by the American Institute of Aeronautics and Astronautics, Inc., with permission.

a nonlinear, nonconvex optimization problem7. Nongradient-based search, such as the genetic algorithm technique, has been utilized in trajectory shaping to compute lookup-table-based line-of-sight bias that is included in a proportional navigation guidance law to enable interception of a target with specific terminal constraints8. Simple on-line guidance laws, such as biased PNG, have been proposed to solve a subset of the aforementioned problem: aiming at satisfying range and terminal impact angle requirements. The so-called time-varying biased PNG is one of them9. Such scheme, designed to achieve zero miss distance and a desired attitude angle of a missile at impact with the addition of a bias term in the acceleration command, is derived under assumptions of planar motion of pursuer and target, idealized autopilot and seeker dynamics (instantaneous response to commands), and relatively slow target maneuvers. Interestingly, under the aforementioned assumptions, guarantees on zero miss distance and a prescribed terminal angle exist. Numerical simulations with realistic 6-DOF engagement models showed that a biased PNG law can steer a rocket, in closed-loop with a classical PI-type autopilot, to a stationary target over a wide range of terminal impact angles within some tolerances10. We propose a trajectory shaping over the entire course of flight obtained by means of a Cross-EntropyMinimization-Based Search (CEMBS). To deal with complex dynamic systems and control constraints, an iterative simulation-based approach is adopted. The Cross-Entropy method has been developed originally for rare-event simulations and to solve complex combinatorial optimization problems such as the Traveling Salesman Problem (TSP)11. The main contribution of this paper lies in the generation of a finite-in-time sequence of lateral acceleration commands obtained with a non-homogeneous Markov chain which exploits the concept of CEMBS. These commands are then utilized to yield a sample set from which a batch of simulations of the guided projectile is performed. Analysis of the resulting statistics drives new iterations by updating the Markov chains transition matrices. The trajectory shaping solution is found when the metrics, defined with respect to the target set, is minimized. The proposed approach is compared to the results obtained by means of exhaustive search. It is shown that a near-optimal trajectory shaping solution can be computed in a relatively short period of time for a given target set, which is the key practical benefit of the proposed trajectory shaping. Importantly, it is shown that the smoothing filter contributes to significantly reducing the energy consumption of the projectile, which prevents the designer from complicating the optimization process by including a term related to the energy expenditure in the objective function or in employing additional constraints.

II. Trajectory Shaping of Guided Projectile

A. Context and Problem Formulation We consider a guided projectile whose nonlinear model is given as ~, w) , & = f (X ,u X

(1) Y = [x y v ]T = g ( X ), T T T where state X=[Xp , Xc ] , results from nonlinear flight dynamics in closed loop with controllers such as a guidance law, an autopilot, and servomechanisms (see Fig. 1). Xp denotes the state-space vector of the projectiles dynamics whereas Xc corresponds to the state-space vector of various control laws. w represents exogenous disturbances such as wind turbulence. The output signals x, y, v, and in (1) stand for the range, the altitude, the speed, and the flightpath angle of the projectile, respectively. Equation (1) includes algebraic nonlinearities that are typically related to flight dynamics and hard nonlinearities such as actuator saturation.

2 American Institute of Aeronautics and Astronautics

Trajectory Shaping
Altitude Latax

Y(t) u u Guidance Law t

Wind Turbulence


w Autopilot Actuators Flight Dynamics of the Projectile INS x y v

& = f ( X , u, w) Closed-Loop Dynamics : X Fig. 1. Trajectory shaping commands and closed-loop dynamics.

For the purpose of designing a CEMBS method for trajectory shaping, a three-degree-of-freedom model is used to represent the flight dynamics of the projectile. As such, the problem is formulated in terms of the lateral acceleration command sets that are applied directly to the closed-loop system shown in Fig. 1. Extension of the method to a 6degree-of-freedom model and waypoint command sets can be readily accomplished. Phases of the guided projectile flight are shown in Fig. 2. During the ascending phase of the ballistic motion, the autopilot aims at de-spinning the projectile. The trajectory shapings output, u, remains zero. Then, the finite-state trajectory shaping is applied from apogee at ta to impact (y=0) at tf. In the remainder of the paper, t * f corresponds to

tf when the trajectory is ballistic.

4000 3500 3000 2500 2000 1500 1000 500 0 -500

Ballistic trajectory

y (m)

Despinned projectile u0

Trajectory of the guided projectile u0

6000 x (m) 8000





Fig. 2. Phases of the guided projectile flight.

Definition 1 (Target Set). The set of constraints that the projectile must satisfy at the end of flight is defined by (2) G={(x,y, v, ) 3[-, 0] / xmin x xmax, v min v v max , min max, y=0}. The proposed trajectory shaper is composed, as shown in Fig. 3, of a finite-state command generator in series with a smoothing filter, the latter having the following dynamics & =F X u filter ( X u , u ), ~ ( X , u ). u =G
filter u

Integrating a smoothing filter in the trajectory shaper allows the derivation of a continuous-time lateral acceleration command whose pattern is expected to better comply with the bandwidth of the projectiles controllers. 3 American Institute of Aeronautics and Astronautics

Augmented Closed-Loop Dynamics Trajectory Shaper u u


Smoothing Filter

Closed-Loop Dynamics

& = f ( X , X u, w)

x v

Fig. 3. Trajectory shaper composed of finite-state command generator and smoothing filter

The smoothing filter can be interpreted as being part of the augmented black-box dynamics shown in Fig. 3 for which the proposed iterative simulation-based search is applied, as explained in Section II.B. The finite-state command generator is defined as follows. Definition 2 (Finite-state command generator). The finite-state command generator provides a piecewise constant signal u(t)=ukU, for all t[tsw,k, tsw,k+1), where U ={-U, -(n-1)U /n, ,-U/n,0, U/n,, (n-1)U /, U}, U, and n. tsw,k , defined for all k=1,,p, are time instants at which switches may occur. tsw,k are fixed prior flight and are uniformly distributed over [ta, tf*], where tsw,1=ta, tsw,p= tf* and [1, 2]; that is, tsw,k- tsw,k-1=( tsw,p - ta)/(p-1) for all k=2,,p. Let z denote the sequence {uk, 1kp}.

Control Objective. Knowing Definitions 1 and 2, find the sequence of control signals z that minimizes dG (xf,vf,f; z) where dG (xf,vf,f; z) = dG (xf; z)+ dG (vf; z)+ dG (f; z), and dG (f; z) =

0 if f [ min max ], min(| min f | / min , | max f | / max ), otherwise,


where f denotes either xf, vf, or f. Subscript f indicates that variables are valued at tf defined by y(tf)=0. Argument

z in dG specifies that the terminal state is obtained by applying the sequence of control signals z to the system in

Remarks. Parameter in Definition 2 is selected such that tsw,n= tf* bounds the time of flight entailed by every possible sequence z. has to be selected carefully since (i) switching time instants are uniformly distributed over [ta, tf*] and (ii) control signals, and thus trajectories, after tf are irrelevant. Consequently, too large a parameter would result in tsw,k >tf for some k<p and thus in unused entries {uk,, up} of z, as illustrated in Fig. 4 with k=p-1.


t sw,1 t sw, 2 t * f

t f > t* f

t sw, p 1 t sw, p = t * f

Flight Time

Fig. 4. Selection of and tsw,k.

4 American Institute of Aeronautics and Astronautics

Fixing tsw,k and forcing z to evolve in the finite set Up enables satisfying the control objective by means of a combinatorial optimization, where one seeks to minimize a metric defined with respect to G by deriving an appropriate sequence z, as explained in Section III. B. Iterative Simulation-Based Approach to Trajectory Shaping The rationale behind the proposed iterative simulation-based approach can be explained as follows. First, a sample of command sequences Z1N is drawn from an independent, identically distributed random sequence Zi associated with zUp. Second, a batch of simulations is carried out for every sequence Zi. Third, the results of the simulations are analyzed in order to drive the selection, at the next iteration, of a new sample set Z1N that tends to decrease dG (xf,vf,f; z). We derive CEMBS in the next section, from which Z1N is obtained. One nave approach would consist of implementing an exhaustive search in lieu of CEMBS. Such a technique becomes prohibitive as p, or n, increases since each zUp corresponds to one of (2n+1)p permutations with repetition. Yet, exhaustive search has been implemented in this study as a yardstick against which one may evaluate the performance of CEMBS. CEMBS, a stochastic search method, is proposed as a means to quickly find a near-optimal solution to the trajectory shaping problem. It is to be noted, however, that the time required to find the solution depends on the computational load associated with solving (1) and on the size of the sample set Z1N . As our ultimate aim is to use a detailed model of the guided projectile, efficiency of the search algorithm is thus critical to achieve fast computation of near-optimal z.

III. Search Based on Cross-Entropy Minimization

A. Review of the Cross-Entropy Method The cross-entropy method11 requires finding a probability distribution that aims at seeking a near-optimal solution z+ in the sense that {z+ Up, dG (xf,vf,f; z+) <+ } is a rare event for some level + typically close to zero; that is,
l = Prh dG ( x f , y f , f ; z + ) < +

(4) = Eh I {d ( x , y , ; z + ) < + }, G f f f is very small. h(Z,P) in (4) denotes a probability distribution of random variable Z that is parametrized by some stochastic matrix P. P is derived in the next section. Searching (z+, +) constitutes the associated stochastic problem (ASP) to that defined in the Control Objective. This problem is related to rare-event simulation since ASP requires estimating l, which is obtained by means of an iterative algorithm that depends on and P11. The guideline that leads to this algorithm is briefly recalled. Instrumental to simulation of rare events is the use of the importance sampling, which provides an unbiased estimator of l
1 l = N

i =1

( x f , y f , f ; Z i ) <

}W ( Z i ) ,


where W ( Z i ) = h(Z,P)/ h( Z ) is the likelihood ratio and (Z1,, Zn) is a random sample drawn from h . By utilizing (5), one seeks to compute l with a sample set that would be smaller than that used by direct application of MonteCarlo estimator of l is h*(Z)= I {dG ( x f , y f , f ; Z i ) < }h( Z , P ) / l as detailed in Ref. [12]. Since l is unknown, computation of l can be achieved simulation to (4). The distribution allows finding the best
h( Z ) that

if one is able to find a family of probability distributions h( Z , P ) along with the reference parameter P such that the cross-entropy between h( Z , P ) and h*(Z) is minimized, yielding the following CEMBS algorithm11.

Step 1. Initialize P to P0. Select (0,1) , (0,1) , and .

5 American Institute of Aeronautics and Astronautics

Step 2. (rth iteration) Proceed as follows as long as the stopping condition in Step 3 is not met. Step 2.1. Draw N samples {Z1 ,, ZN}= Z1N according to h( Z , Pr 1 ) . Step 2.2. Select the

N best draws {Z1,..., Z N } as defined by dG (xf,vf,f; z), which is obtained by

simulation of (1) over [ta, tf*]. Let r be the N -th order-statistics of the sequence {dG (xf,vf,f; Z 'i ), i=1,,

N }.

,..., Z N , the reference parameter Q that minimizes the cross-entropy, which is Step 2.3. Update, from Z1

obtained by maximizing

ln h' (Z i, Q ) .
i =1

Step 2.4. Update Pr as follows: Pr = Q + (1 ) Pr 1 Step 3. (Stopping condition) r == r +1 or there exists i=1,,N such that dG (xf,vf,f; Zi)=0.
B. Proposed Reference Parameter The cross-entropy method has been successfully applied to various combinatorial optimization problems among which TSP is solved by selecting h( Z , P ) as a function of a homogeneous Markov chain that expresses the probability of being at a given site at tk conditioned to the location of the salesman at tk-1 11. We consider here a nonhomogeneous Markov chain13, defined for all k{2,..,p} by
Pk = pk , ij

[ ]1i, j 2n +1,


pk ,ij = Pr(U k = u k | U k 1 = u k 1 ),

which models transition from uk-1U at tsw,k-1 to ukU at tsw,k. Notice that the transition matrix in (6) depends on k. A homogeneous Markov chain may be too restrictive when applied to a dynamical system with terminal state constraints since a k-independent transition matrix would force every transition to be governed by the same probabilities. The sample Zi is thus selected at each iteration of the cross-entropy algorithm from the set of p-1 matrices {P2,,Pp}. The first entry of Zi is randomly selected from U with a uniform distribution. The probability of generation of zUp is given by
h(z,P) = pk ,ij I {zZij ( k )} ,
k = 2 i =1 j =1 p 2 n +12 n +1


where Zij (k ) denotes the set of z such that uk-1=i and uk=j. Step 2.3 of CEMBS involves minimizing of the cross-entropy between h( Z , P ) and h*(Z). This minimization along with the use of Lagrangian multipliers ui,k, which ensure that the ith row of pk,ij sums to one, is equivalent to the following maximization over Pk 0
p 2 n +1 2 n +1 max min E Pk I {dG ( x f , y f , f ;Z )< } ln h ( Z , P ) + u i ,k p k ,ij 1 . j =1 ui , k Pk k = 2 i =1

(8) Expressing ln h( Z , P) as
ln h(z,P) =

k = 2 i =1 j =1

I {zZ (k )} ln pk ,ij ,

p 2 n +12 n +1


and differentiating (8) with respect to pk,ij yields 6 American Institute of Aeronautics and Astronautics

E Pk I {dG ( x f , y f , f ; Z )< }I {Z Zij ( k )} pk ,ij

2 n +1

+ ui,k = 0 ,


Summing (10) over j=1,,2n+1, where obtains

pk ,ij =

pk ,ij = 1 , and letting

j =1

Zi (k ) denote the set of z such that uk-1=i, one

E Pk I {dG ( x f , y f , f ; Z ) < }I {Z Zij ( k )} E Pk I {dG ( x f , y f , f ; Z )< }I {Z Zi ( k )}


The estimator of (11) where expectations are replaced by averaging over the sample Z1N ={Z1 ,, ZN} is expressed as

q =1 I {d ( x , y , ;Z )< }I {Z Z (k )} . k ,ij = p N q =1 I {d ( x , y , ;Z )< }I {Z Z (k )}

G f f f q q ij G f f f q q


s ijth entry is given by p ,..., P . P k ,ij in The reference parameter Q in Step 2.3 of CEMBS is thus equal to P k 2 p
(12), where is replaced by r obtained in Step 2.2. Determining the estimator in (12) involves the computation of (p-1)(2n+1)2 parameters. Thus, the number of levels in U and of switching times tsw,k, which can be interpreted as the quantization and the time discretization of uk with respect to a continuous-time trajectory shaping signal, respectively, have a significant impact on the computation time required to find a near-optimal sequence z. While one may be tempted to increase p and n so that z approaches an analog signal, the limited available computational resources constrain the resolution of z. Furthermore, the dynamics of projectile acts as a filter, which actually bounds from above practical values of p.

IV. Simulations
Applying Newtons law of motion to a projectile and assuming that the drag force predominates over the lift force yields the following model of a de-spinned projectile uv * & x = Cd (v x w x ) y , v (v , y )v mv (13) uv y * & y = Cd (v, y )v (v y w y ) + g, v mv & = vx , x & = vy , y
v ~ . vs ( y ) , m, S, , and u 2m vs ( y ) stand for the sound speed, which is a nonlinear function of the altitude, the mass of the projectile, the reference area, the air density, and the control input signal, respectively. A typical drag coefficient Cd ( M ) profile is shown in Fig. 5.
2 * = (v x w x ) 2 + ( v y w y ) 2 , C d + v2 where v = v x ( v, y ) = y , v

SCd ( M )

, and M =

7 American Institute of Aeronautics and Astronautics




0.2 0.15 0.1 0




Mach Number
Fig. 5. Drag coefficient model.

~ in (13) is expressed as the filtered and delayed command u To represent neglected dynamics, the input signal u

2 e d s ~ ( s )) , sat10 N (u (14) 2 2 2 s + s + where function sat10 N(u) saturates the lateral acceleration at 10N. The following parameter values are used: m=18.5 kg, S=18.810-3 m2, and g=9.81 m/s2. A ballistic trajectory is obtained with the following initial conditions: x(0) =0 m, y(0) =0 m, v(0) = 625 m/s, and (0)=/4 rad. The apogee [6271, 3804] m is reached at 24.7 s. The range of the course is 11276 m. The impact of the projectile on the ground occurs at 54.4 s with a terminal velocity of v*(tf)=268 m/s and terminal flight-path angle of -1.05 rad (-60.2). The target set is defined by v min = 241.6 m/s, v max= 1.1v*(tf), min=-1.15 rad, max=-0.94, xmin=10599 m, and xmax=10609 m, which imply that a precision of 10 m (relative error of 9.4310-2 %) has to be met. The algorithm proposed in Section III is implemented with n=1, p=15, U=92.5 N, =1.1, =0.08, =0.1, and N=p2. The closed-loop systems in 2 2 /( s 2 + 2ns + n )u ' ( s ) , with n= 2 (1) is augmented with the following second-order smoothing filter, u ( s ) = n rad/s and =0.9, which is inserted in series with (13) and (14). The stochasticity of the cross-entropy method may entail fluctuating performances from one simulation to another in terms of computation time and distance dG (xf,vf,f; z) at impact. One hundred and fifty simulations were run on a 1Gb, 2.4-GHz Xeon computer with Matlab/Simulink-compiled code14 in order to analyze the performance of the proposed approach. CEMBS is tested with p{12,13,14,15,18,20}. The empirical frequency that dG (xf,vf,f; z)=0 is depicted in Fig. 6. The average relative error, which is computed when dG (xf,vf,f; z)0, is shown Fig. 7. The relative error is defined as min(|xmin-x|/xmin, |x-xmax|/xmax).
u (s) =

The average computation time is about 2 min 30 s when p{12,13,14,15} and increases to 3 min 48 s and to 5 min 30 s when p=18 and p=20, respectively. Note that an exhaustive search with p=10 entailed 310 = 59049 executions and lead to a total computation time of 4 h 30 min. For this case, exhaustive search has not resulted in dG (xf,vf,f;

z)=0. Increasing the number of switches improves, as expected, the frequency of obtaining a trajectory that reaches the target set, although the curves slope in Fig. 6 tends to decrease as p is increased. The same applies to the average relative error. The stochastic nature of the cross-entropy-based search incurs near optimality rather than optimality and thus necessitates trading off between fixing a desirable confidence level that the target set is reached and bounding from above the computation time. For the cases studied, p=15 appeared to be a good compromise.

8 American Institute of Aeronautics and Astronautics

100 95 Empirical Frequency of Zero Terminal Error (%) 90 85 80 75 70 65 60 55 50 10



16 18 Number of Switches



Fig. 6. Empirical frequency of dG (xf,vf,f; z)=0.

5 4.5 4 Average Terminal Error (%) 3.5 3 2.5 2 1.5 1 0.5 0 10



16 18 Number of Switches



Fig. 7. Average terminal error when the target set is not reached.

The projectiles trajectory is shown in Figs. 8 and 10 with smoothing filters defined by n= 2 rad/s and =0.9, and by n= 1 rad/s and =0.9, respectively. It can be seen that the resulting command is smooth compared to the square-wave command sequence, as shown in Fig. 8(b), while still leading to the prescribed target set. The energy expenditure measure of the lateral acceleration command, u2dt, is shown in Figs. 9 and 11. This measure is computed for the filtered and unfiltered command signals. As expected the filter is instrumental to obtaining various profiles of smoothed acceleration commands that better comply with the bandwidth limits of the autopilot and the actuators.

9 American Institute of Aeronautics and Astronautics

4000 3500 3000 2500 2000 1500 1000

100 80 60 40 u, u (N)
0 2000 4000 6000 x (m) 8000 10000 12000

20 0 -20 -40

y (m)


0 -500

-80 -100 0 5 10 15 Time (s) 20 25 30



Fig. 8. Trajectory shaping with smoothing filter n= 2 rad/s and =0.9 obtained with CEMBS. (a) Trajectories of guided projectile (---) and ballistic flight (); (b) Sequence of commands z.

2 1.8 1.6 1.4 1.2

x 10

Unfiltered command Actual command


0.8 0.6 0.4 0.2 0 0 5 10 15 Time (s) 20 25 30

Fig. 9. Energy expenditure measure, u2dt, for unfiltered (---) and actual () commands with filter n= 2 rad/s and =0.9

10 American Institute of Aeronautics and Astronautics

4000 3500 3000 2500 2000 1500 1000

100 80 60 40

u, u (N)
0 2000 4000 6000 x (m) 8000 10000 12000

20 0 -20 -40

y (m)

500 0 -500

-60 -80 -100


15 Time (s)






Fig. 10. Trajectory shaping with smoothing filter n= 1 rad/s and =0.9 obtained with CEMBS. (a) Trajectories of guided projectile (---) and ballistic flight (); (b) Sequence of commands z.

18 16 14 12 10 8 6 4 2 0

x 10

Unfiltered command Actual command


15 Time (s)




Fig. 11. Energy expenditure u2dt for unfiltered (---) and actual () commands with filter n= 1 rad/s and =0.9. Table 1. Comparison of energy expenditure with and without filter n2/(s2+2ns+n2). MEAN VALUE REDUCTION OF ENERGY u2dt CONSUMPTION TRAJECTORY SHAPER WITHOUTFILTER TRAJECTORY SHAPER

1.56105 1.19105 8.51104 7.82104

n= 2 rad/s; =0.9 n= 1 rad/s; =0.9 n= 1 rad/s; =1.2

23 % 45 % 50 %

to the same value. As reported in Table 1, the presence of a smoothing filter entails a decrease (23%) of the average energy consumption, which is significant enough to justify the inclusion of such filter into the trajectory shaping function. Similarly to the choice of p, which, owing to the dynamics of projectile, is bounded from above, 11 American Institute of Aeronautics and Astronautics

To evaluate the gain in energy savings, 100 simulations were run with the guided projectile. In those runs, trajectory shaping is carried out with and without the smoothing filter. The results are reported in Table 1 for several values of the filter parameters. It is clear that the presence of a filter can be used as a simple and convenient means to reduce the maneuvering effort (u2dt) of the projectile, while maintaining empirical frequency of dG (xf,vf,f; z)=0

the choice of the filters bandwidth is dictated by the bandwidth of the projectiles dynamics. Indeed, decreasing the natural frequency from 2 rad/s to 1 rad/s while maintaining the damping ratio to 1 allows to further reduce the energy consumption.

V. Conclusion
An iterative simulation-based trajectory shaping scheme is proposed so that a sequence of commands permits a guided projectile to be steered toward a prescribed target set. The sequence is obtained by means of an appropriately defined Markov chain whose transition matrices are updated by minimizing a metric between the projectiles terminal state and the prescribed target set. The search is based on cross-entropy minimization and achieves computation of near-optimal sequence of commands in a relatively short period of time when compared to an equivalent solution found by the exhaustive search method. Furthermore, including a smoothing filter constitutes a practical means of bounding the energy expenditure of the guided missile without resorting to more complex energy-constrained objective functions.

I.M. Mitchell, A.M. Bayen, and C.J. Tomlin, Time-Dependent Hamiltonian-Jacobi Formulation of Reachable Sets for Continuous Dynamics Games, IEEE Transactions on Automatic Control, Vol. 50, No. 7, 2005, pp. 947-957. 2 A.M. Bayen, I.M. Mitchell, M.M.K. Oishi, and C.J. Tomlin, Aircraft Autolander Safety Analysis Through Optimal ControlBased Reach Set Computation, Journal of Guidance, Control, and Dynamics, Vol. 30, No. 1, 2007, pp. 68-77. 3 M.M.K. Oishi, I.M. Mitchell, C.J. Tomlin, and P. Saint-Pierre, Computing Viable Sets and Reachable Sets to Design Feedback Linearizing Control Laws under Saturation, Proceedings of the 45th IEEE Conference on Decision and Control, San Diego, CA, USA, 2006. 4 J. Aubin, Viability Theory, Birkhauser, 1991. 5 E. Cruck and P. Saint-Pierre, Nonlinear Impulse Target Problems Under State Constraints: A Numerical Analysis Based on Viability Theory, Set-Valued Analysis, Vol. 12, N. 4, 2004, pp. 383-416. 6 M. Fliess, J. Lvine, P. Martin, and P. Rouchon, Flatness and Defect of Nonlinear Systems: Introduction Theory and Application, International Journal of Control, No. 61, 1995, pp. 1327-1361. 7 M.B. Milam, K. Mushambi, and R.M. Murray, A New Computational Approach to Real-Time Trajectory Generation for Constrained Mechanical Systems, Proceedings of the39th IEEE Conference on Decision and Control, Sidney, Australia, 2000. 8 H. Brown Cribbs, Genetics-Based Trajectory Discovery for Tactical Missiles, AIAA 1st Intelligent Systems Technical Conference, AIAA Paper 2004-6550, September 2004. 9 B.S. Kim, J.G. Lee, and H.S. Han, Biased PNG Law for Impact with Angular Constraint, IEEE Transactions on Aerospace and Electronic Systems, Vol. 34, No. 1, 1998, 277-288. 10 C.A. Rabbath and R. Lestage, Biased proportional navigation guidance law - Concept and preliminary results, DRDC Valcartier TM 2010-306, Nov. 2010. 11 R.Y. Rubinstein and D.P. Kroese, The Cross-Entropy Method - A unified Approach to Combinatorial Optimization, MonteCarlo Simulation and Machine, Information Science and Statistics, New York, NY: Springer, 2004. 12 R.Y. Rubinstein and B. Melamed, Modern Simulation and Modeling, New York, NY: Wiley Series in Probability and Statistics, 1998. 13 P. Brmaud, Markov Chains, Gibbs Fields, Monte Carlo Simulation, and Queues, New York, NY: Springer-Verlag, 2001. 14 Simulink 7 User's Guide, The MathWorks, Inc., Natick, MA, USA, 2007.

12 American Institute of Aeronautics and Astronautics