I.

INTRODUCTION

Three-Dimensional Midcourse Guidance Using Neural Networks for Interception of Ballistic Targets

EUN-JUNG SONG MIN-JEA TAHK Korea Advanced Institute of Science and Technology

A suboptimal midcourse guidance law is obtained for interception of free-fall targets in the three-dimensional (3D) space. Neural networks are used to approximate the optimal feedback strategy suitable for real-time implementation. The fact that the optimal trajectory in the 3D space does not deviate much from a vertical plane justifies the use of the two-dimensional (2D) neural network method previously studied. To regulate the lateral errors in the missile motion produced by the prediction error of the intercept point, the method of feedback linearization is employed. Computer simulations confirm the superiority of the proposed scheme over linear quadratic regulator (LQR) guidance and proportional navigation (PN) guidance as well as its approximating capability of the optimal trajectory in the 3D space.

Manuscript received May 5, 1999; revised May 9, 2001; released for publication December 11, 2001. IEEE Log No. T-AES/38/2/11431. Refereeing of this contribution was handled by T. F. Roome. Authors current addresses: E-J. Song, School of Mechanical and Aerospace Engineering, Seoul National University, Seoul, Korea; M. J. Tahk, Division of Aerospace Engineering, Dept. of Mechanical Engineering, Korea Advanced Institute of Science and Technology, 373-1, Kusong Yusong Taejon 305-701, South Korea, E-mail: ([email protected]).

c 2002 IEEE 0018-9251/02/$17.00 404

For most ground-based defense systems being considered today, an essential element of midcourse guidance is to guide the missile with the maximum probability of collision. Midcourse guidance is often formulated as an optimal control problem to shape the trajectory to maximize the terminal energy or to minimize the flight time. The implementation of the optimal midcourse guidance law is not easy since a nonlinear two-point boundary value problem is to be solved to obtain the optimal trajectory. Direct numerical solutions to this problem introduces a heavy in-flight computational burden and the convergence characteristics may not be acceptable. Solving the problem in real time is often not feasible. Furthermore, a feedback guidance law is not readily provided from the numerical solution. Several techniques such as a perturbation procedure [13], linear quadratic regulator (LQR) with a database of the optimal trajectories [4, 5], and modified proportional guidance [6] have been previously proposed for the real-time implementation of optimal midcourse guidance. Recently, the approximation ability of the artificial neural network [7] is used to derive an on-board guidance algorithm suitable for real-time implementation [8]. The key idea of this approach is to train a neural network to extract the functional relationship between optimal commands and missile states from the set of the optimal trajectories computed in advance. Then, the trained network constitutes a feedback guidance law, which allows the missile to adapt perturbations in the target states and its own trajectory as well as to follow the optimal trajectory. For implementation, only the weights of the network are needed to be stored. The neural-network approach has been extended for two-dimensional (2D) midcourse guidance against moving targets [9]. For intercept point prediction, a time-to-go estimator using neural networks has been devised to consider the time-varying characteristics of the missile velocity. Also, a more elaborate neural-network scheme has been derived to improve the robustness to variations in the launch condition [10]. For this purpose, a new input pattern that is robust to the launch condition has been employed. In this article, the neural-network approach is extended for three-dimensional (3D) midcourse guidance of a missile system to intercept nonmaneuvering targets decelerated by atmospheric drag. Although many applications of neural networks on missile guidance and control have appeared with growing interest [1113], no application to the 3D midcourse guidance problem has been attempted. Direct extension of the previous neural-network method to the 3D midcourse guidance problem is not plausible since the number of data to be learned

IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 38, NO. 2 APRIL 2002

by the neural network is tremendously larger than the 2D case. The neural network should also have a very complicated structure. To avoid this difficulty, we propose a guidance law for which the neural-network approach is used for the guidance in the vertical plane but the feedback linearization technique [14] for lateral control. This approach is motivated by the fact that the optimal trajectory is confined to a vertical plane if the correct intercept point is known and the missile is launched vertically or toward the intercept point. Hence, if the error in prediction of the intercept point is corrected during the initial phase of the flight, then the proposed 3D guidance method is expected to work well. To predict the intercept point accurately, we need to precisely compute the time to go of the missile. For the prediction of the missiles flight time, we use an additional neural network that learns the time-to-go characteristics from the optimal trajectory data. Since the target is supposed to be intercepted at a high altitude, it is reasonable to assume that the target motion is affected only by the gravity forces. Hence, the target trajectory is a Keplerian orbit and the future position can be computed without direct integration of the equation of motion. In this study, a trade-off study with other existent methods such as LQR with a database of optimal trajectories and proportional navigation (PN) guidance is performed to demonstrate the effectiveness of the proposed neural-network approach. The performance of each method is evaluated in terms of memory requirements, optimality, and intercept performance. This paper is organized as follows. The mathematical model of a typical medium-range surface-to-air missile is introduced first. The formulation and development of the midcourse guidance law using neural networks are then described, and the prediction algorithm with the time-to-go estimator is explained. Next, simulation results are presented to investigate the performance of the proposed scheme. Finally, the conclusions of this work are summarized. II. MATHEMATICAL MODEL For problem formulation, the Earth-centered Earth-fixed (ECEF) frame and the navigation north, east, down (NED) frame are used [15]. The interrelationship between the coordinate systems and the state variables are illustrated in Fig. 1, where denotes the Earth rotational speed. The Earth is assumed to be a spherically symmetric body. The state variables are the missile position in the ECEF frame (r, , ), the missile velocity relative to the ECEF frame v, and the flight-path angle and heading angle , respectively. The control variables are the angle of attack and the angle of total lift direction , called the bank angle (the missile need not be a bank-to-turn

Fig. 1. Geometry of coordinate frames (x, y , z : inertial frame, xe , ye , ze : ECEF).

type). The equations of motion are given by _ = v sin r v cos sin r cos _ = v cos cos r (T cos D ) _= g sin v m _= + r 2 (cos2 sin cos sin cos cos ) _ = (T sin + L) sin + v sin cos sin mv cos r cos + r 2 sin cos sin 2 cos sin cos v cos cos (5) (1) (2) (3)

(4)

+ 2 sin _= (T sin + L) cos g cos v cos + mv v r + r 2 (cos2 cos + sin cos sin cos ) v

+ 2 sin cos where

2 L= 1 2 v SCL , 2 D= 1 2 v SCD ,

(6)

CL = CL ( 0 )
2 CD = CD0 + kCL :

The aerodynamic derivatives CL , CD0 , and k are given as functions of Mach number M , which is a function of v and the altitude h: CL = CL (M ), CD0 = CD0 (M ), k = k (M ):

For the spherical Earth model, the expression for g is as follows: 2 Re g = g0 Re + h where g0 is the gravity magnitude at the surface of the Earth and Re is an effective radius of the Earth. The missile mass and thrust are given as functions of time t: m = m(t), T = T (t ):
405

SONG & TAHK: THREE-DIMENSIONAL MIDCOURSE GUIDANCE USING NEURAL NETWORKS

The position and velocity of the target are provided by a ground support system to the missile for computation of the intercept point. This study considers a ballistic target that is only affected by the central gravitational field. III. REAL-TIME SUBOPTIMAL MIDCOURSE GUIDANCE For long-range and medium-range missiles, the optimal trajectory shaping ensures an extended range and more favorable endgame conditions. However, its direct formulation based on optimal control theory results in a two-point boundary value problem, for which the computational time is too long for real-time implementation. To solve this problem, a suboptimal guidance method using neural-network approximation has been suggested in [8, 11]. A neural network is trained to learn the closed-loop optimal command in terms of the current missile states and terminal conditions, (7) u (t) = g (x(t), xf ) using the optimal trajectory data generated off-line. In [11], a terminal guidance law is synthesized to minimize the kill vehicles divert propellent. They show that the terminal guidance law implemented by a neural network can increase the engagement range. The feasibility of the neural-network approximation has also been investigated for a midcourse guidance problem constrained in the vertical plane [9]. Modifications of this work have also been proposed in [10] to provide robustness against variations in the missile launch condition. One of the most important aspects of neural-network design is how to construct the training data [16]. Reference [10] shows that reorganization of the input patterns reduces the sensitivity of the neural-network guidance scheme to the trajectory perturbations produced by the _ -feedback launch error. In this study, we use the guidance law proposed in [10], for which the optimal angle-of-attack command is assumed to be _ rather dependent on the line-of-sight (LOS) rate than the flight-path angle ; where _= _ , x xf , h hf ) = (v, (h hf )v cos + (x xf )v sin : (x xf )2 + (h hf )2 (8)
Fig. 2. Definition of guidance plane.

On the other hand, if the effect of Earth rotation is neglected and the missile is launched vertically, the optimal missile motion is confined within a vertical plane determined by the missiles initial position and intercept point, denoted as the guidance plane in Fig. 2. Hence, if prediction of the intercept point is accurate, the optimal 2D missile motion in the guidance plane can approximate the optimal 3D motion. The 3D guidance command is then decomposed into two commands; one to track the optimal trajectory in the guidance plane and another to handle the lateral motion of the missile not to deviate from this plane. The neural-network guidance law of (8) is used to implement the former while feedback linearization is employed for the latter. A. 3D Guidance Law The 3D guidance law generates commands for the angle of attack and the bank angle . Using the neural-network guidance law of (8), is commanded as q N xN )2 + (xE xE )2 , xD xD ) _ , (xI ( v , I M M I M (11)
N E D N E D where (xI , xI , xI ) and (xM , xM , xM ) are the predicted intercept point and current missile position in the NED frame defined at the launch position of the missile, respectively. The role of the bank-angle command is to steer the missile to the direction of the predicted intercept point given by E E xI xM = tan1 : (12) N xN xI M

(9)

A straightforward way to extend the 2D neural-network guidance method to the 3D guidance problem is to train a neural network to learn the 3D optimal trajectory data. However, this requires a large amount of training data as well as a complicated structure of the neural network: _ yaw , x xf , y yf , h hf ): _ pitch , = g(v, (10)
406

The dynamics in (5) is rewritten as _ = (T sin + L) sin + mv cos (13)

where represents the last four terms of the right-hand side (RHS) of (5). These terms are produced by the rotation and roundedness of the Earth. Inspection of (13) shows that we can choose

IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 38, NO. 2 APRIL 2002

Fig. 4. Intercept geometry.

should be commanded as ( ) _ yaw mNv 1 = sin : T sin + L

(20)

Fig. 3. Guidance law implementation for intercept in 3D space. (a) Neural-network guidance. (b) LQR with database of optimal trajectories.

B. Intercept Point Prediction Since the target is assumed a ballistic object falling in the central gravitational field, the target trajectory is Keplerian; the initial condition and the gravitational field determine the trajectory. Above an altitude of approximately 60 km, aerodynamic loads are much smaller than the gravitational forces [18]. Consequently, a fairly adequate representation of the target trajectory can be obtained by ignoring the aerodynamic forces entirely. A missile-target intercept geometry in the 3D space is depicted in Fig. 4, where A denotes the current target position, B the current missile position, and I the predicted intercept point. The target trajectory from A to I is a Keplerian orbit satisfying rT I r ( ) = 1 cos cos(T + ) + , cos2 T cos T

as = sin
1

mv cos (ulateral ) , T sin + L

jj

2 (14)

to cancel the nonlinear terms, and to linearize the map from ulateral to , which is described by _ =u lateral : (15)

Thus, the tracking control problem can be handled by using linear control theory. If ulateral is chosen as ulateral = kc ( ) (16) then the proper choice of the parameter kc enables the missile to maintain . Here, kc is chosen as inverse of the time constant of the closed-loop dynamics: _ = k ( ): c (17)

2 vT = =rT

(21)

In this study, the command of (13) is implemented without . Note that can be neglected in the medium-range missile case since it is sufficiently small when compared with the first term in (13). The proposed guidance law shown in Fig. 3(a) consists of neural-network guidance for the vertical missile motion and a controller for lateral control with an additional block for prediction of the intercept point. Conventional guidance laws can also be employed for the lateral command. For pure PN guidance [17], the bank-angle command is designed as follows _ yaw ayawc = Nv (18)

where ()T denotes the target states, ()I the intercept point, and the central angle, respectively. I and I are computed by rotating ~ rT around ~ rT ~ vT by . Therefore, given ~ rT and ~ vT , there is a unique relationship between ~ rI and . Also, the time from A to I is given by [18]
T tgo =

vT cos T f(2 )(1 cos )=( cos2 T ) + cos(T + )= cos T g + 2rT vT (2= 1) tan1 3=2

rT ftan T (1 cos ) + (1 ) sin g

(2= 1)1=2 : cos T cot(=2) sin T

(22)

_ yaw where N is the effective navigation constant and is the yaw component of the LOS rate of the predicted intercept point with respect to the missile body axes. Then, from 1 ayawc = (T sin + L) sin (19) m

M (time For the missile, the rough approximation of tgo to go from B to I ) by Range=v is not appropriate since there are significant changes in the missile velocity during the midcourse guidance phase. Here, an additional neural network is employed to estimate M tgo as proposed in [9]. This neural network is trained M -function of the optimal trajectory, to learn the tgo

SONG & TAHK: THREE-DIMENSIONAL MIDCOURSE GUIDANCE USING NEURAL NETWORKS

407

which is assumed to be
M tgo

with the dynamic constraints (23) _ = A x + B u x @f @f = x + u @ x (t,x,u) @ u (t,x,u)

M tgo (v, , x xf , h hf ):

Assume that the deviation of from is sufficiently M small, then tgo in the 3D space can be approximated M by the tgo estimator valid for 2D vertical-plane guidance
M M N N E E D D = tgo (v, , , xI xM , xI xM , xI xM ) tgo q M N x N ) 2 + ( xE xE ) 2 , xD x D ) : (v, , (xI tgo I M M I M

(29)

_ = f (t, x, u) is the nonlinear missile dynamics. where x Then, the guidance law is given as u = K (t) x = R 1 B T P x where P satisfies the time-varying matrix Ricatti equation _ = PA AT P Q + PBR 1 B T P , P P (tf ) = S: (31) (30)

(24)
T M Both tgo and tgo are given as functions of and I () is given by the root of the equation M T () = tgo ( ): tgo

(25)

The Earth rotation neglected in the above formulation produces some errors. But the accuracy in prediction of the intercept point improves gradually as the missile approaches the target. C. Trade-Off Study Neural-network guidance is compared with two existing approaches. One is LQR based on a database of the optimal trajectories (LQR guidance) [5], and the other is proportional navigation guidance (PN guidance) which is a conventional guidance law used frequently. The implementation issue of each method is briefly discussed in the following. 1) Neural-Network Guidance: The angle-of-attack command in (11) is implemented by the multilayer feedforward networks: !!! X X X
= glinear wk gsigmoid wkj gsigmoid wji zi
k j i

Since the in-flight computation of (31) is difficult, the precomputed gain K (t) and the corresponding reference trajectory (t, x, u) are stored prior to launch. Therefore, its implementation requires a larger memory size than that of neural-network guidance which requires only the weights of neural networks. Fig. 3(b) shows the block diagram of the LQR guidance implementation. The efficient construction of the database is another problem [20]. 3) PN Guidance: The pitch and yaw-axes acceleration commands are calculated by _ pitch + g cos apitchc = Npitch v _ yaw : ayawc = Nyaw v (32) (33)

It is simpler than neural-network guidance but the nonoptimality during the midcourse guidance phase degrades its intercept performance as shown in the numerical results. IV. NUMERICAL RESULTS A. Neural-Network Design For the missile model illustrated in Fig. 5, the neural-network guidance law and the tgo -estimator are designed. The optimal control problem is defined to minimize the performance index J = tf subject to inequality constraints j(t)j 5 , ((t) = 0, 0 t 57 s t > 57 s): (35) (34)

(26) where g () is an activation function, wkj is the synaptic weight from the neuron j to the neuron k , and q zi is the ith element of the input vector

N xN )2 + (xE xE )2 xD xD 1]. Whenever _ (xI [v I M M I M the input vector is applied to the sensory nodes of the network, the command is obtained by propagating it through the network, layer by layer. 2) LQR Guidance: The command of this guidance law is a combination of a nominal command u obtained from a database of optimal trajectories and a real-time correction command u to correct perturbations from the reference trajectory;

u = u + u : u is calculated by solving the linear quadratic problem defined as [19] Z tf S x + ( xT Q x + uT R u) dt J = xT f f

t0

(27)

(28)

Then, it is converted to a parameter optimization problem [21], and solved by the sequential quadratic programming (SQP) method [22]. Our conversion method is direct shooting. Once the initial guesses of the discretized control history (t) and the final time tf are given, the state differential equations are integrated explicitly using the 4th-order RungeKutta

408

IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 38, NO. 2 APRIL 2002

Fig. 5. Missile data. (a) Mass and thrust profile. (b) Aerodynamic derivatives.

approach. Terminal conditions are chosen as the set of 9 points

(xf , hf ) = f(40, 40), (40, 60), (40, 80), (60, 40), (60, 60), (60, 80), (80, 40), (80, 60), (80, 80)g (km)

in the vertical plane. The same vertical launch condition 0 = 90 , v0 = 27 m/s, (x0 , h0 ) = (0, 0) km (36)

Fig. 6 shows the 9 optimal flight trajectories used for the training of the neural networks, for which the procedure is identical to that described in [8]. The dotted lines represent the expected region of target intercept. The structure of the neural network for vertical-plane guidance is 2 hidden layers, with 7 and 6 neurons in each layer, repectively, and the tgo -estimator is composed of the same number of hidden layers with 5 and 4 units in each layer. B. Performance Evaluation The proposed neural-network guidance law combined with the tgo -estimator, shown in Fig. 3(a),
409

is applied to all cases. The rotation of the Earth is neglected here so that the problem is a 2D one.

SONG & TAHK: THREE-DIMENSIONAL MIDCOURSE GUIDANCE USING NEURAL NETWORKS

Fig. 6. Optimal trajectory data used for neural-network training. TABLE I Simulation Results Without Prediction Errors Optimal tf (sec) 58.70 59.58 66.46 NN Guidance (3D) tf (sec) (error %) MD (m) (tgo ) (sec) 58.73 (0.05) 59.59 (0.02) 66.55 (0.14) 270.34 330.51 433.03 0.15 0.13 0.21

Target Case 1 Case 2 Case 3

Note: MD : miss distance. (tgo ) = 1=tf average estimation error of tgo .

R tf
0

jtgo true tgo estimated jdt:

is tested by computer simulation. Depending on the initial position and velocity of the target, three different scenarios illustrated in Fig. 7 are considered in the 3D space. The 3D midcourse guidance law is applied until the time of intercept without a terminal homing phase. The feedback gain for correction is chosen as kc = 0:4 and the predicted intercept point is updated at every 5 s. For comparision, the optimal trajectory to the scenario is also calculated by using the SQP method. 1) Neural-Network Guidance: Table I shows simulation results for the case of no initial prediction errors. The launch direction of the missile is determined by the predicted intercept point. It shows that the performance of the 3D guidance law is comparable to that of the optimal trajectory. The miss distances obtained here can be easily reduced if homing guidance starts several kilometers away from the intercept point. In the second performance test described in Table II(A), the effect of initial prediction errors is considered; the initial heading angle of the missile is commanded toward the target, not the predicted
410

Fig. 7. Target initial conditions.

intercept point. The 2D guidance law is again applied to intercept a fictitious target fixed at the terminal target position obtained by the simulation results of the 3D guidance law, as shown in Fig. 6. In this case, the rotation of the Earth is ignored. In spite of the prediction error and the effect of Earth rotation, the guidance performance is not much degraded; specifically, the increase in flight time does not exceed 0.14%. This table also shows that the performance of the 3D guidance law is not much different from the ideal 2D guidance case. For medium range missiles, the rotation of the Earth does not produce a significant effect on the missile trajectory. Although the choice of kc = 0:4 satisfies the suboptimal performance, a higher value (kc = 5:0) during the initial flight can reduce the flight time

IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 38, NO. 2 APRIL 2002

Fig. 8. Simulation results of case 1. (a) Flight trajectory. (b) Time to go (sec). (c) Azimuth angle.

TABLE II Simulation Results With Prediction Errors A. kc = 0:4 target Case 1 Case 2 Case 3 NN Guidance (3D) tf (sec) (error %) MD (m) 58.74 (0.07) 59.59 (0.02) 66.55 (0.14) 403.05 336.94 447.31 (tgo ) (sec) 0.16 0.13 0.20 tf (sec) 58.73 59.60 66.58 NN Guidance (2D) MD (m) (tgo ) (sec) 193.84 193.69 81.43 0.15 0.15 0.23

B. kc = 5:0 target Case 1 Case 2 Case 3 NN Guidance (3D) tf (sec) (error %) MD (m) (tgo ) (sec) 58.73 (0.05) 59.59 (0.02) 66.55 (0.14) 282.22 330.09 434.57 0.14 0.13 0.21

further by a small amount, as can be seen in Tables II(A) and II(B). Note that these results are close to the result obtained for the case of no prediction errors. However, prediction errors can induce saturation of the bank-angle command so that kc should be carefully chosen by considering various intercept scenarios.
411

SONG & TAHK: THREE-DIMENSIONAL MIDCOURSE GUIDANCE USING NEURAL NETWORKS

Fig. 8. (Continued.) (d) Bank angle. (e) Angle of attack. (f) Velocity. (g) Flight-path angle. (h) Flight trajectory in east-height plane. (i) Flight trajectory in east-north plane.

Fig. 8 shows the simulation results of Case 1 for kc = 0:4. In Fig. 8(a), the discrepancy from the optimal flight trajectory is too small to be observed.
412

Fig. 8(b) shows that the predicted time to go of the missile coincides well with the true time to go. The direction of the predicted intercept point is also

IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 38, NO. 2 APRIL 2002

TABLE III Simulation Results of Other Guidance Laws LQR Guidance tf (sec) (error %) MD (m) 59.00 (0.51) 59.73 (0.25) 67.94 (2.23) 209.57 475.51 3218.82 PN Guidance tf (sec) (error %) 59.64 (1.60) 60.58 (1.68) 67.59 (1.70)

Target Case 1 Case 2 Case 3

(tgo ) (sec) 0.12 0.02 1.44

MD (m) 1338.27 2386.79 19931.55

close to the heading angle of the optimal trajectory as shown in Fig. 8(c). It takes about 10 s for the missile to achieve its heading in the direction of . On the other hand, the choice of kc = 5:0 makes the missile achieve the commanded heading more quickly. It accompanies larger bank-angle commands than the case of kc = 0:4 during the initial period as shown in Fig. 8(d). The angle of attack, velocity, and flight-path angle are shown in Figs. 8(e)8(g), respectively. It is observed that the 3D guidance case is quite close to the optimal trajectory as well as the ideal 2D guidance so that the proposed guidance law can be effectively used for the 3D midcourse guidance problems. 2) LQR and PN Guidance: A massive trade-off study between neural-network guidance and two existent approaches described in Section III is conducted. The optimal trajectories used to obtain neural-network guidance are chosen as nominal trajectories for LQR guidance. Although nominal trajectories should be selected carefully to cover the entire missile flight envelope, for the purpose of comparison it is reasonable to assume common trajectories are available for both guidance laws. For each trajectory, the optimal gain matrix is computed by backward integration of the matrix Riccati equations. The time to go of the missile is also calculated by interpolating the flight-time data of the optimal trajectories as briefly illustrated in Fig. 3(b). The simulation results are presented in Table III. Case 1 and Case 2 show performance comparable to the optimal but their flight times are slightly longer than those of neural-network guidance. It results from the fact that the LQR correction command to compensate the off-nominal condition is different from the real optimal command for the same condition. The time histories of the angle of attack, velocity, flight-path angle, and vertical flight trajectory in Figs. 8(e)8(h) show some differences between LQR guidance and the optimal solution. As the flight condition deviates more from the nominal value, nonoptimality increases, resulting in degraded performance (see Case 3 in Fig. 6). Therefore, apart from the advantage of simple implementation, the neural-network approach is superior to LQR guidance in terms of ability to deal with off-nominal (not-trained) flight conditions. In a similar manner, PN guidance is applied against the terminal target position of the 3D neural-network guidance law. The performance evaluated against the fixed target is expected to be

similar to that against the predicted intercept point whose variation is so small, as shown in Fig. 8(a). Both N in (32) and (33) are set to 5. As shown in Table III, PN guidance produces larger miss distances and longer flight times in all cases than both of LQR guidance and neural-network guidance. From Figs. 8(e) and 8(h), we see that the angle of attack during the initial period of the flight is too small to reach the altitude of the target. On the other hand, Fig. 8(i) shows that the lateral channel has no problem. It again confirms our claim that the synthesis of the optimal command is not necessary in the lateral channel while its employment is important in the vertical channel. V. CONCLUSION

The neural-network guidance method, which has been previously studied for real-time midcourse guidance in the 2D space, is extended to the case of 3D flight for interception of nonmaneuvering ballistic targets. By decomposing the 3D guidance problem into the 2D guidance problem in the vertical plane and the lateral control problem, a computationally efficient guidance law is obtained. For the vertical motion, a neural network is trained to produce the optimal angle-of-attack command for tracking of the projection of the optimal trajectory onto the vertical guidance plane. For lateral control, the method of feedback linearization is used to steer the missile toward the direction of the predicted intercept point. An additional neural network is employed as a tgo -estimator to predict the intercept point. Computer simulations show that the proposed guidance law generates a good approximation of the 3D optimal trajectory while providing better performance than the existing methods such as LQR guidance and PN guidance.
REFERENCES [1] Cheng, V. H. L., and Gupta, N. K. (1986) Advanced midcourse guidance for air-to-air missiles. Journal of Guidance, Control, and Dynamics, 9, 2 (1986), 135142. Menon, P. K. A., and Briggs, M. M. (1990) Near-optimal midcourse guidance for air-to-air missiles. Journal of Guidance, Control, and Dynamics, 13, 4 (1990), 596602. Dougherty, J. J., and Speyer, J. L. (1997) Near-optimal guidance law for ballistic missile interception. Journal of Guidance, Control, and Dynamics, 20, 2 (1997), 355362. 413

[2]

[3]

SONG & TAHK: THREE-DIMENSIONAL MIDCOURSE GUIDANCE USING NEURAL NETWORKS

[4]

[5]

[6]

[7]

[8]

[9]

[10]

[11]

Imado, F., Kuroda, T., and Miwa, S. (1990) Optimal midcourse guidance for medium-range air-to-air missiles. Journal of Guidance, Control, and Dynamics, 13, 4 (1990), 603608. Imado, F., and Kuroda, T. (1992) Optimal missile guidance system against a hypersonic target. AIAA Paper 92-4531. Newman, B. (1996) Strategic intercept midcourse guidance using modified zero effort miss steering. Journal of Guidance, Control, and Dynamics, 19, 1 (1996), 107112. Hornik, K., Stinchcombe, M., and White, H. (1990) Multilayer feedforward networks are universal approximators. Neural Networks, 2 (1990), 359366. Song, E. J., Lee, H., and Tahk, M. J. (1996) On-line suboptimal midcourse guidance using neural networks. In Proceedings of the 35th SICE Annual Conference, Tottori University, Japan, 1996, 13131318. Song, E. J., and Tahk, M. J. (1998) Real-time midcourse guidance with intercept point prediction. Control Engineering Practice, 6, 8 (1998), 957967. Song, E. J., and Tahk, M. J. (1999) Real-time midcourse missile guidance robust to launch conditions. Control Engineering Practice, 7, 4 (1999), 507515. Cottrell, R. G., Vincent, T. L., and Sadati, S. H. (1996) Minimizing interceptor size using neural networks for terminal guidance law synthesis. Journal of Guidance, Control, and Dynamics, 19, 3 (1996), 557562.

[12]

[13]

[14] [15] [16] [17]

[18] [19] [20]

[21]

[22]

Fu, L. C., Chang, W. D., Yang, J. H., and Kuo, T. S. (1997) Adaptive robust bank-to-turn missile autopilot design using neural networks. Journal of Guidance, Control, and Dynamics, 20, 2 (1997), 346354. Geng, Z. J., and McCullough, C. L. (1997) Missile control using fuzzy cerebellar model arithmetic computer neural networks. Journal of Guidance, Control, and Dynamics, 20, 3 (1997), 557565. Khalil, H. K. (1996) Nonlinear Systems. Englewood Cliffs, NJ: Prentice-Hall, 1996, 8185. Siouris, G. M. (1993) Aerospace Avionics Systems: A Modern Synthesis. San Diego, CA: Academic Press, 1993. Zurada, J. M. (1992) Introduction to Artificial Neural Systems. St. Paul, MN: West Publishing, 1992, 214216. Ghose, D. (1996) Lecture Notes on Guidance Laws and Their Applications. Automatic Control Research Center, Seoul, Korea, 1996, 90115. Regan, F. J., and Anandarkrishnan, S. M. (1993) Dynamics of atmospheric re-entry. AIAA, Washington, DC, 1993. Bryson, A. E., Jr., and Ho, Y. C. (1975) Applied Optimal Control. Halsted Press, 1975, 146174. Hardtla, J. W., and Milligan, K. H. (1987) Design and implementation of a missile guidance law derived from modern control theory. AIAA paper 87-2447. Hull, D. G. (1997) Conversion of optimal control problems into parameter optimization problems. Journal of Guidance, Control, and Dynamics, 20, 1 (1997), 5760. Lawrence, C., Zhou, J. L., and Tits, A. L. (1996) Users Guide for CFSQP Version 2.4: A C Code for Solving (Large Scale) Constrained Nonlinear (Minimax) Optimization Problems, Generating Iterates Satisfying All Inequality Constraints. TR-94-16rl, Institute for Systems Research, University of Maryland, College Park, MD, 1996.

Eun-Jung Song received the M.S. degree in 1997 and the Ph.D. degree in 2000, both in aerospace engineering, from Korea Advanced Institute of Science and Technology. She visited University of California at Los Angeles as a postdoctoral researcher from 2000 to 2001. Currently, she is a postdoctoral researcher of the School of Mechanical and Aerospace Engineering, Seoul National University, Korea. Her major research interests include neural networks, missile guidance, and aircraft navigation.

Min-Jea Tahk received the B.S. degree from Seoul National University, Seoul, Korea, and the M.S. and Ph.D. degrees from the University of Texas at Austin, in 1983 and 1986, all in aerospace engineering. From 1976 to 1981 he was a research engineer at the Agency for Defense Development, and from 1986 to 1989 he was employed by Integrated Systems, Inc., Santa Clara, CA. He is presently Associate Professor of Aerospace Engineering at Korea Advanced Institute of Science and Technology, Taejon, Korea. His research interests include target tracking, missile guidance, flight control, and evolutionary computation.
414 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 38, NO. 2 APRIL 2002