Professional Documents
Culture Documents
158
Artificial Neural Networks in Microwa¨ e CAD 159
example consider a filter response in terms of its ent sources as, e.g., from experimental tests andr
geometrical parameter.. or different numerical methods.
Nowadays, the relevant and rapidly evolving The rapidly evolving field of ANN applications
market of microwave wireless communication is in microwave CAD and optimization has wit-
determining a paradigm shift in the design of nessed several excellent contributions: different
microwave components: not only the electrical problems have been successfully attacked; new
characteristics being of relevance, but also other methodologies have been introduced and signifi-
issues such as reduced time to market Ži.e., fast cant progress has been made in this dynamic area
design., yield optimization, manufactured-ori- of microwave CAD. It seems therefore appropri-
ented design, tolerance analysis, etc., are becom- ate to try to summarize in this review paper the
ing increasingly important. In other words the progress made so far. The structure of the work is
development of microwave components is moving the following: in Section II the necessary theoret-
from ‘‘performance-oriented’’ design to ‘‘statisti- ical background on neural networks is provided;
cal’’ computer-aided design ŽCAD. w1x. Unfortu- then, in Section III, we illustrate the current
nately, to perform statistical CAD with current applications of ANN in microwave CAD and, in
approaches does not appear feasible: a single Section IV, we discuss some typical issues in
analysis of a component may often require sev- ANN implementation. In Section V we summa-
eral hours while, in practice, hundreds of analysis rize the research trends and innovations; in par-
are required. As a possible remedy, sophisticated ticular, we also introduce the use of self-organiz-
techniques, like the Adjoint Network Method w2᎐6x, ing maps in the microwave context. Finally, in
and Space Mapping w7᎐10x have been recently Section VI, we speculate on some possible future
introduced. However, both techniques still make applications of ANN in microwave CAD.
use of computer-intensive electromagnetic ŽEM.
full-wave simulators while all is needed is a multi-
dimensional nonlinear approximator which re- II. NEURAL NETWORKS:
lates the input parameters to the output ones: in A BRIEF DESCRIPTION
this respect artificial neural networks ŽANNs. ap-
pear to be a good candidate. In fact, as noted in Several concepts in neurocomputing have been
IEEE Proceedings special issue on ANN w11, pp. inspired by studies of biological networks. From
1353᎐1576x, ‘‘While in the 1980s many of the one perspective neural networks can be thus con-
applications proposed produced results which sidered as a way to further understanding brain
were barely competitive with existing approaches, functions; from a different standing point, they
the 1990s have revealed a variety of application can be regarded as an extension of the conven-
areas where neural networks are becoming one of tional data processing techniques developed over
the tools which provide the best results, both several decades. This latter perspective is the one
regarding quality of the outcome and the ease of we consider: since, as engineers, we are mostly
implementation.’’ interested in understanding how neural networks
It is therefore not surprising that ANN are compare with different processing techniques for
currently finding useful applications also in the problem solving.
microwave context; in fact, to cite again w11, pp. Neural networks are massively parallel, highly
1353᎐1576x: ‘‘ . . . they are especially useful in situ- connected structures consisting of a number of
ations where a model-based or parametric ap- simple, nonlinear processing elements; because of
proach to information processing is difficult to their massively parallel structure, they can per-
formulate.’’ Actually, for microwave problems, the form computations at a very high rate if imple-
model-based approach is indeed feasible but, of- mented on a dedicated hardware; because of their
tentimes, it is rather inconvenient with respect to adaptive nature, they can learn the characteristics
the computer resources involved. For the latter of input signals and adapt to changes in the data;
problems ANN provide an excellent complement because of their nonlinear nature they can per-
to existing models and codes. In fact, in several form functional approximation and signal filtering
cases, existing codes can be used in order to train operations which are beyond optimal linear tech-
the ANN and to provide some a priori knowledge niques.
when needed; moreover ANN can accumulate Several neural paradigms are available in the
experience and merge data obtained from differ- technical literature; all of them, in some sense,
160 Burrascano, Fiori, and Mongiardo
derive from the pioneering work of McCulloch B. Feed-Forward Layered Networks and
and Pitts w12x, Hebb w13x, and Rosenblatt w14x, Error Backpropagation
whose work introduced the model of neural net-
Typically feed-forward layered networks consist
works as computing machines, the basic model of
of a set of source nodes which constitute the
self-organization, and the model of learning with
input layer, one or more hidden layers of compu-
a teacher, respectively. During the 1960s and
tation nodes and an output layer of computation
1970s the work of a number of researchers
ŽAmari, Fukushima, Grossberg, Kohonen, Wi- nodes. Each processing node Žneuron. performs a
weighted sum of the signal components at its
drow. prepared the start over of the field with the
input; this sum is thus fed into a block performing
publication of the works of Hopfield w15x, which
a differentiable nonlinear processing Žusually a
related symmetric dynamical networks to opti-
sigmoidal logistic function. of the type,
mization, and Rumelhart, Hinton, and Williams
w16x, which introduced the method of backpropa-
gation. 1
f Ž¨ . s . Ž2.
In the following, a concise description is given 1 q exp Ž y¨ .
of three neural paradigms Žthe Hopfield network,
the multilayer perceptron and Kohonen self- The input signal propagates through the net-
organizing maps., which have been considered for work in a forward direction, on a layer-by-layer
microwave problems. basis. These networks are usually referred to as
multilayer perceptrons ŽMLPs.. The processing
task to be performed is described by a set of
A. The Hopfield Recurrent Model input᎐output data Žthe training set.: the network
The Hopfield model relies on the energy function, specializes to solve this task by modifying its
parameters Žweights. by means of an iterative
optimization procedure Žthe learning procedure..
1
Esy The error backpropagation, described in w16x, is a
2
Ý wi j S i S j , Ž1.
widely popular learning technique that provides
i/j
an efficient and algorithms stable way to correct
the network weights.
where
from the distribution of the input signals: each sentation formula can be realized by a three-layer
cluster of frequently occurring input stimuli is feed-forward layered network. In 1989 several
represented by a large area in the map. papers related to this topic were published
w19᎐21x: the authors of these papers, following
different techniques, claimed that three-layer
D. Function Approximation by Learning feed-forward layered networks with sigmoid units
in the hidden layer can approximate continuous
One of the most widely considered applications of
or other kinds of functions defined on compact
neural networks in the field of microwave CAD is
sets in R n. Their results can be summarized as
for functional approximation. The capability to
follows Žuniversal approximation theorem.: Let
approximate a multivariate, nonlinear function is
Ž⭈. be a nonconstant, bounded, and monotone-
a powerful ability of neural models which is out-
increasing continuous function. Let l p denote
lined in the following.
the p-dimensional unit hypercube w0, 1x p. The
Approximation theory deals with the problem
space of continuous functions on l p is denoted
of approximating or interpolating a continuous,
by C Ž l p .. Then, given any function f g C Ž l p .
multivariate function f Ž x . by an approximating
and ⑀ ) 0, there exist an integer M and sets of
function F Ž w, x . having a fixed number of param-
real constants ␣ i , i , and wi j , where i s 1, . . . , M
eters w belonging to some set P. For a choice of
and j s 1, . . . , p such that we may define
a specific F, the problem is then to find the set of
parameters w that provides the best possible ap- M p
proximation for f on the set of the available
‘‘examples’’ Žlearning step.. It is fundamental to
F Ž x1 , x 2 , . . . , x p . s Ý ␣i žÝ wi j x j y i ,
/
is1 js1
choose an approximating function F that can Ž3.
represent f as well as possible. The problem of
learning a mapping between an input and output
space is equivalent to the problem of synthesizing as an approximate realization of the function f Ž⭈.;
an associative memory that retrieves an appropri- that is
ate output when presented with the input, and
generalizes when presented with new inputs. It is < F Ž x 1 , x 2 , . . . , x p . y f Ž x 1 , x 2 , . . . , x p . < - ⑀ , Ž4.
also equivalent to the problem of estimating the
system that transforms inputs into outputs given a for all x 1 , x 2 , . . . , x p 4 g l p . The coefficient ⑀
set of examples of input᎐output pairs: according somehow measures the degree of approximation
to this interpretations both unsupervised and su- accuracy.
pervised neural networks have been extensively This theorem is directly applicable to MLPs
studied as functional approximators. Žsee Fig. 1.: we first note that the logistic function
The problem of approximating a function of is a nonconstant, bounded, and monotone-in-
several variables by MLPs has been studied by creasing continuous function; it therefore satisfies
many authors: in w17x Wieland and Leighten deal the conditions imposed on Ž⭈.. Then we note
with the capabilities of networks consisting of one that Ž3. represents the output of a MLP consist-
or two hidden layers; Irie and Miyake w18x ob- ing of p input nodes and a single hidden layer
tained an integral representation formula with an of M neurons with thresholds i , and input-to-
integrable kernel defined beforehand: this repre- hidden weights wi j ; the output neuron is a linear
available which accurately model passive 䢇 VLSI interconnect models, i.e., transmission
structures into the millimeter-wave fre- line analysis with nonlinear loads have been
quency range. EM simulation effectively simulated by using ANN in w43᎐47x; as noted
models passive element dispersion and mu- in w47x ‘‘Modeling of lossy interconnects is
tual coupling effects ignored by traditional done by electromagnetic ŽEM. simulation
circuit simulation tools. However, EM simu- techniques, which involve the numerical
lation methods take tremendous computa- simulation of Maxwell’s equations or vari-
tional efforts and are not practical for inter- ants thereof. Full-wave three-dimensional
active CAD . . . The execution time for a Ž3D. EM analysis, being approximation free,
full-wave EM simulation depends on the gives very accurate results, but is highly
number of grid points used when meshing CPU-intensive, and thus is not feasible for
the structure and the numerical tolerances. on-line use in large scale CAD and opti-
The number of grid points greatly depends mization techniques.’’ For this type of appli-
on the smallest critical geometric feature.
cation ANN can provide substantial on-line
For example, the time required for the EM
speed-up with respect to conventional meth-
simulation of a 2.5-turn spiral inductor on a
ods Žas an example consider Table IV of w47x
5- m grid is about 3 min per frequency
where EM simulation takes 20᎐80 h, while
point. However, the time required for a sim-
ANN takes about 40᎐130 s. In the same
ilar inductor on a 2- m grid is about 53 min
per frequency point. These simulation times work, data reported in Table V show that
clearly limit the practicality of EM simula- NILT Žnumerical inversion of Laplace trans-
tion in an interactive CAD environment.’’ form. takes about 34.43 h, AWE Žasymptotic
waveform evaluation. takes 9.56 h and the
ANN uses only 6.67 min..
Moreover, since EM tools are also used in
䢇 Yield optimization of X-band amplifiers has
order to generate the training set, the above
statements clearly show why some form of design been considered as an example in w43x. In
of experiment ŽDoE. is absolutely necessary for this case a three-stage X-band MMIC am-
these cases. plifier was considered and the Khatibzadeh
and Trew model has been compared with a
䢇 Large signal HEMT characterization w42x neural model; the speed-up ratio using the
and physics-oriented FET model w43᎐46x. In neural model was 6 for the optimization
w42x the bias-dependent behavior of the part, 30 for Monte Carlo analysis, and 5 for
HEMT has been characterized in terms of yield optimization Žsee Table VI of w43x..
the intrinsic elements of a conventional 䢇 Tolerance analysis of waveguide filters has
small-signal equivalent circuit. Noticeably, been studied in w48x. In this case it was
the intrinsic element data depending on two noted that ‘‘after the learning phase the
input parameters, Vg s and Vd s , have been trained network gives its response in a few
obtained from the S-parameters measure- milliseconds and allows to approximate any
ments performed at various bias settings. filter response in the range of geometrical
Also in the same work it has been found values we considered Ž40 m for each di-
that a fi¨ e-layered configuration composed mension, 4 = the nominal tolerances.
of only 28 neurons is adequate to simultane- . . . The computational cost associated to the
ously represent seven bias-dependent intrin- modal analysis procedure is over 100= more
sic elements from Vg s and Vd s . In w44x expensive than the one associated to the
instead, a three-layer feed-forward neural neural model.’’
network has been considered for modeling a
FET having six input parameters Žgate-
length L, gate width W, channel thickness
a, doping density Nd , and Vg s and Vd s .; 100 IV. TYPICAL ISSUES IN ARTIFICIAL
neurons were used in the hidden layer. NEURAL NETWORK IMPLEMENTATION
However, in this case an ingenious proce-
dure making use of the sparsity of the inter- Naturally, except for w33, 34x Žwhich address a
nal activations of neural networks has been fairly different problem., some common topics
devised. are almost invariably present when applying ANN
Artificial Neural Networks in Microwa¨ e CAD 165
the most important topics to consider in this batch mode ranged from 20 s to 9 min.’’ Another
context, and which will be discussed in detail in important point to note concerns the generaliza-
the following subsections, are: tion capability: according to w46x ‘‘Moving to the
extrapolation region, the accuracy of KBNNs de-
䢇 knowledge-based ANN teriorates much more slowly than that of MLPs.’’
䢇 use of sparsity To summarize, the main advantages of knowl-
䢇 selection of the training set via DoE edge-based ANN are:
䢇 accuracy improvement by self-organizing
maps ŽSOM.; i. the possibility of using already developed
equivalent circuits or methods of analysis;
A. Knowledge-Based Artificial ii. the dimensionality reduction of the training
Neural Network set and the consequent reduction of the
training time;
It is well known that a large arsenal of models
iii. the improved generalization capabilities.
already exist for most of the cases considered in
microwave CAD, e.g., equivalent circuits devel-
It seems fair to conclude that the pioneering work
oped for microstrip via holes or for CPW discon-
of the above authors have paved the way to very
tinuities. Two approaches have been reported so
exciting developments for the application of neu-
far in the ANN microwave literature in order to
ral networks together with the methods᎐
take advantage of previous knowledge: the hybrid
models᎐equivalent circuits already in use in mi-
EM-ANN modeling w37x and the knowledge-based
crowave CAD and, in Section VI, we therefore try
neural models w46x.
to draw some possible future developments in this
According to w37x: ‘‘The hybrid EM-ANN model
area.
is formed by generating the difference in S-
parameters between the existing approximate
model and the EM simulation results Ž ⌬ S .. The B. Selection of the Training Set Via
Ž ⌬ S . data is then used to train the EM-ANN Design of Experiment
model.’’ The main advantage is that a simpler Another way to reduce the dimension of the
input᎐output relationship is considered, hence training set is to carefully select the training set
requiring less EM simulation points. In particu- points according to concepts on the design of
lar, for the two-port broadband GaAs microstrip experiment introduced by Taguchi. The applica-
via example considered in w37x the number of EM bility of these concepts to microwave CAD has
simulations points needed to train the hybrid been recently illustrated in w1x. In ANN modeling
model for a given accuracy has been almost half of microwave problems the DoE has been used
that needed for the complete model. Errors of typically in order to reduce the amount of time-
the same order have been obtained by using 15 consuming EM full-wave simulations, e.g., in
training points for the hybrid model and 29 train- w36᎐39x; while fractional-factorial experimental
ing points for the complete model. Accordingly, design has been adopted in w40, 41x.
also the training time has been significantly re- It seems likely that, since the above investiga-
duced Ž5 h 44 min for the hybrid model; 11 h 35 tions have demonstrated the usefulness of DoE in
min for the complete model; times are relative to ANN, future works may increasingly rely upon
an HP 700 workstation.. these types of techniques.
In w45, 46x a knowledge-based neural network
ŽKBNN. has been introduced. As noted in w46x,
C. Use of Sparsity
same work, after extensive comparison in several
cases, ‘‘ . . . the overall tendency suggests that the The sparsity concept is commonly used in the
accuracy of KBNN trained by a small set of circuit analysis area and in microwave CAD w49x
training data is comparable to that of MLP trained and, in w44x it has been introduced in the ANN as
by a larger set of training data.’’ As far as the a new form of training, the sparse-training tech-
training time is concerned, according to the au- nique. The basic remark for introducing this tech-
thors, ‘‘The CPU time for MLP training by the nique is that for each sample many neuron activa-
conventional sample-by-sample error backpropa- tions are close to zero; this gives rise to sparsity in
gation approach ranged from 22 to 60 min. The the network and has no effect on weights update.
CPU time for MLP or KBNN training by pro- Naturally the speed-up achievable with this tech-
posed gradient based l 2 optimization approach in nique depends on the sparsity present: reduction
168 Burrascano, Fiori, and Mongiardo
of CPU time of more than 50% have been re- ing a number of ‘‘classes’’ of filter behaviors such
ported in w44x. that the mean square difference between all
curves within the same class is minimized. Model-
D. Self-Organizing Maps ing the filter curves within each one of these
clusters significantly reduces the problems evi-
In w48x it has been shown that a feed-forward
denced above. To obtain this result two main
neural model can represent the amplitude re-
steps have to be performed: the first one is to
sponse of a microwave filter even if the tolerances
define an automatic procedure to cluster similar
taken into account Ž"40 m. are much wider
than those which have to be expected in a usual curve behaviors, the second one is to define a
manufacturing process Ž"10 m.. It was also neural model for each cluster of filter responses.
noted, moreover, that an increase in the tolerance The first step was performed by training a
range implies a significant wider dispersion of Kohonen ‘‘self-organizing neural map’’ ŽSOM.
filter response behaviors Ždecreased accuracy.. To w50x. In particular, a SOM was trained with as
enhance the accuracy of the model its topology many input nodes as the number of frequency
has to be significantly increased. Unfortunately, samples considered to define each filter response.
training a neural model of increased topological The number of neurons of the SOM equals the
complexity leads to an excessive computational number of classes we want to identify in the set of
effort required during the training phase. filter responses. Figure 5 shows the results ob-
In order to define a general neural modeling tained in the case of four classes considered.
procedure, we can introduce a scheme which Once the group of filter responses belonging to
overcomes this problem by automatically identify- each class is defined, a MLP is trained on the
Figure 5. When using SOM a number of classes of filter behaviors are identified such that
the mean square difference between all curves within the same class is minimized. For the
waveguide filter case, a SOM with four neurons has produced the above four classes.
Artificial Neural Networks in Microwa¨ e CAD 169
responses of each class. In this way a feed-for- tional effort is reduced because:
ward neural model is obtained for each one of the
classes. 䢇 training a SOM is an extremely light task.
After the SOM and the MPLs have been 䢇 Training each MLP implies a relatively light
trained the overall modeling procedure can be computational cost because it models a set
performed, as described in Figure 6. of homogeneous filter responses.
䢇 We start by giving the geometrical parame- In particular, in the experiments whose results
ters to a first step MLP, which gives an early are shown in Figure 7, the overall training time
approximation of a filter response. In this was about 80% than that required to train the
case a relatively simple topology can be used. single step MLP model.
The results reported refer to a 13-35-1 net-
work Ž35 neurons in the hidden layer., much
lighter to be trained than the 13-50-1 topol- VI. POSSIBLE FUTURE
ogy adopted in the case of single step model. DEVELOPMENTS
䢇 This first step approximation is used to feed
We will conclude our review with the challenging
the SOM, whose output indicates the class
task of suggesting some possible future develop-
which the filter response belongs to.
ments of ANN in microwave CAD. The chance of
䢇 The third and final step is then to select the writing a dreambook is naturally present but,
neural model associated to that specific class: nonetheless, we consider it worthwhile to under-
its output gives the required filter response. take this risk.
ample, the drain current of a FET as a function crowave models are available but they are
of its gate length, gate width, channel thickness, too slow for design purposes;
doping density, gate voltage, and drain voltage.. 䢇 since significant knowledge has already been
However, the knowledge neurons may also be developed the use of knowledge-based neu-
sophisticated CAD tools. As an example, for EM ral models is particularly relevant;
simulation problems, we can imagine knowledge- 䢇 ANN are going to play an important role in
based neurons implementing modal techniques, the most CPU intensive part of microwave
or FEM. Naturally the response of these neurons CAD, namely: yield optimization, tolerance
may require a significant amount of time. There- analysis, manufacturing-oriented design;
fore they are activated only when no other faster 䢇 some forms of design of experiment ŽDoE.
analyses are feasible. In this way the range of will probably become a common tool also in
applicability of the ANN will be increased. microwave CAD in order to reduce the di-
mensions of the training and test sets;
䢇 further improvement of the neural modeling
D. Artificial Neural Network for
Method Hybridization accuracy by means of a preliminary cluster-
ing of the responses via self-organizing maps
ANN may also be useful to perform method seems also to be very promising. This proce-
hybridization at the data level: let us consider, for dure implies that the overall modeling task
example, a waveguide filter which may Žor may is divided into several single modeling pro-
not. present rounded corners or tuning screws, cesses, allowing a higher degree of accuracy
etc. If the filter structure is simple enough then without implying an unacceptable increase
relatively fast modal techniques can be used to of the computational cost associated to the
generate the training set; however when complex training phase.
artifacts like screws are present CPU intensive
FEM or FDTD need to be used for generating It seems fair to conclude that ANN applications
the relative data. In the same way the ANN can in microwave CAD appear to be very promising;
make use of data generated by different numeri- ANN will probably be used not as an alternati¨ e
cal approaches, or even mix them with experi- to existing model-based CAD tools but rather as
mental data. a useful and necessary complement.
5. F. Alessandri, M. Dionigi, M. Mongiardo, and R. 22. C. Stone, Optimal global rates of convergence for
Sorrentino, Fullwave optimization and sensitivity of non-parametric regression, Ann Stat 10 Ž1982.,
waveguide networks: a comparison of approaches, 1040᎐1053.
Int J MIMICAE Ž1997.. 23. H. Akaike, A new look at the statistical model
6. F. Alessandri, M. Mongiardo, and R. Sorrentino, identification, IEEE Trans Automat Contr 19
New efficient full wave optimization of microwave Ž1974., 716᎐723.
circuits by the adjoint network method, IEEE Mi- 24. J. Rissanen, Stochastic complexity and modelling,
crowave Guided Wave Lett 3 Ž1993., 414᎐416. Ann Statist 14 Ž1986., 1080᎐1100.
7. J. Bandler, R. Biernacki, S. Chen, R. Hemmers, 25. D. Rumelhart, G. Hinton, and R.J. Williams, ‘‘The
and K. Madsen, Electromagnetic optimization ex- effective number of parameters: an analysis of gen-
ploiting aggressive space mapping, IEEE Trans Mi- eralisation and regularization in nonlinear learning
crowave Theory Tech 43 Ž1995., 2874᎐2882. systems,’’ Advances in neural information process-
8. J. Bandler, R. Biernacki, and S. Chen, Fully auto- ing systems, J. Moody, S. Hanson, and R. Lipp-
mated space mapping optimization of 3D struc- mann ŽEditors., Morgan Kaufmann, Cambridge,
tures, Int Microwave Symp Dig, 1996, pp. 753᎐756. MA, 1992, pp. 318᎐362.
9. J. Bandler, R. Biernacki, S. Chen, and Y. Huang, 26. S. Amari, N. Murata, and S. Yoshizawa, A criterion
Design optimization of interdigital filters using ag- for determining the number of parameters in an
gressive space mapping and decomposition, IEEE artificial neural model, Proc Int Conf Artificial
Trans Microwave Theory Tech 45 Ž1997.. Neural Networks 1991, pp. 9᎐14.
10. J. Bandler, R. Biernacki, S. Chen, and D. Omer- 27. D.B. Fogel, An information criterion for optimal
agic, Space mapping optimization of waveguide fil- neural network selection, IEEE Trans Neural Net-
ters using finite element and mode-matching elec- works 2 Ž1991., 490᎐497.
tromagnetic simulators, Int Microwave Symp Dig, 28. N. Murata, S. Yoshizawa, and S. Amari, Network
1997. information criterionᎏdetermining the number of
11. Artificial neural networks applications, IEEE Proc hidden units for an artificial neural network, METR
84 Ž1996., special issue. 92-05, Technical Report, University of Tokyo, 1992.
29. K. Yoshitomi, A. Ishimaru, J. Hwang, and J.S.
12. W.S. McCulloch and W. Pitts, A logical calculus of
Chen, Surface roughness determination using spec-
the ideas immanent in nervous activity, Bull Math
tral correlations of scattered intensities and an
Biophys 5 Ž1943., 115᎐133.
artificial neural network technique, IEEE Trans
13. D.O. Hebb, The organization of behavior, Wiley,
Antennas Propagat 41 Ž1993., 498᎐502.
New York, 1970.
30. H. Southall, J.A. Simmers, and T. O’Donnell, Di-
14. F. Rosenblatt, The perceptron: a probabilistic
rection finding in phased arrays with a neural net-
model for information storage and organization in
work beamformer, IEEE Trans Antennas Propagat
the brain, Psychol Rev Ž1958., 386᎐408.
43 Ž1995., 1369᎐1374.
15. J. Hopfield, Neural networks and physical system 31. G. Washington, Aperture antenna shape prediction
with emergent collective computational properties, by feedforward neural networks, IEEE Trans An-
Proc Natl Acad Sci USA Ž1982., 2554᎐2558. tennas Propagat 45 Ž1997., 683᎐688.
16. D. Rumelhart, G. Hinton, and R.J. Williams, 32. A. Patnaik, R.K. Mishra, G.K. Patra, and S.K.
‘‘Learning internal representations by error propa- Dash, An artificial neural network model for effec-
gation,’’ Parallel distributed processing: Explo- tive dielectric constant of microstrip line, IEEE
rations in the microstructure of cognition, Ch. 8. Trans Antennas Propagat 45 Ž1997., 1697.
Foundations, Vol. 1, MIT Press, Cambridge, MA, 33. M. Vai and S. Prasad, Automatic impedance
1986, pp. 318᎐362. matching with a neural network, IEEE Microwave
17. A. Wieland and R. Leighten, Geometric analysis of Guided Wave Lett 3 Ž1993., 353᎐354.
neural network capacity, Proc IEEE first ICNN, 34. M. Vai and S. Prasad, Microwave circuit analysis
Vol. 6, 1987, pp. 385᎐392. and design by a massively distributed computing
18. B. Irie and S. Miyake, Capacity of three layered network, IEEE Trans Microwave Theory Tech 43
perceptrons, Proc IEEE ICNN 1, Vol. 7, 1988, pp. Ž1995., 1087᎐1094.
641᎐648. 35. T. Horng, C. Wang, and N.G. Alexopoulos, Mi-
19. S.M. Carrol and B.W. Dickinson, Construction of crostrip circuit design using neural networks, Int
neural nets using radon transform, Proc IJCNN, Microwave Symp Dig, 1993, pp. 413᎐416.
Vol. 1, 1989, pp. 607᎐611. 36. P. Watson and K. Gupta, EM-ANN models for via
20. K. Hornik, M. Stichcombe, and H. White, Multi- interconnects in microstrip circuits, Int Microwave
layer feedforward networks are universal approxi- Symp Dig, 1996, pp. 1819᎐1822.
mators, Neural Networks 2 Ž1989., 359᎐366. 37. P. Watson and K. Gupta, EM-ANN models for
21. G. Cybenko, Approximation by superpositions of a microstrip vias and interconnects in multilayer cir-
sigmoidal function, Math Contr Signals Syst 2 cuits, IEEE Trans Microwave Theory Tech 44
Ž1989., 303᎐314. Ž1996., 2495᎐2503.
Artificial Neural Networks in Microwa¨ e CAD 173
38. P. Watson and K. Gupta, EM-ANN modeling and 45. F. Wang and Q. Zhang, Knowledge based neural
optimal chamfering of 90( cpw bends with air- models for microwave design, IEEE Int Microwave
bridges, Int Microwave Symp Dig, 1997. Symp Dig, 1997.
39. P. Watson and K. Gupta, Design and optimization 46. F. Wang and Q. Zhang, Knowledge based neural
of cpw circuits using EM-ANN models for CPW models for microwave design, IEEE Trans Mi-
components, IEEE Trans Microwave Theory Tech crowave Theory Tech 45 Ž1997., 2333᎐2343.
45 Ž1997., 2515᎐2523.
40. G.L. Creech, B. Paul, C. Lesniak, T. Jenkins, R. 47. A. Veluswami, M.S. Nakhla, and Q.J. Zhang, The
Lee, and M. Calcatera, Artificial neural networks application of neural networks to em-based simula-
for accurate microwave CAD applications, IEEE tion and optimization of interconnects in high-speed
Int Microwave Symp Dig, 1996, pp. 733᎐736. VLSI circuits, IEEE Trans Microwave Theory Tech
41. G.L. Creech, B. Paul, C. Lesniak, T. Jenkins, and 45 Ž1997., 712᎐723.
M. Calcatera, Artificial neural networks for fast 48. P. Burrascano, M. Dionigi, C. Fancelli, and M.
and accurate EM-CAD of microwave circuits, IEEE Mongiardo, A neural network model for CAD and
Trans Microwave Theory Tech 45 Ž1997., 794᎐802. optimization of microwave filters, Int Microwave
42. K. Shirakawa, M. Shimiz, N. Okubo, and Y. Daido, Symp Dig, 1998, pp. 13᎐16.
A large signal characterization of an HEMT using 49. A. Caproni, F. Cervelli, M. Mongiardo, L. Tarri-
a multilayered neural network, IEEE Trans Mi- cone, and F. Malucelli, Bandwidth reduced full-
crowave Theory Tech 45 Ž1997., 1630᎐1633. wave simulation of planar microstrip circuits, ACES
43. A.H. Zaabab, Q.J. Zhang, and M.S. Nakhla, Neural
Special Issue, 13 Ž1998., 197᎐204.
network modeling approach to circuit optimization
and statistical design, IEEE Trans Microwave The- 50. T. Kohonen, Clustering, taxonomy, and topological
ory Tech 43 Ž1995., 1349᎐1358. maps of patterns, Proc of the VI Int Conf on
44. A.H. Zaabab, Q.J. Zhang, and M.S. Nakhla, Device Pattern Recognition, 1982.
and circuit-level modeling using neural networks 51. P. Watson, K. Gupta, and R.L. Mahajan, Develop-
with faster training based on network sparsity, ment of knowledge based artificial neural network
IEEE Trans Microwave Theory Tech 45 Ž1997., models for microwave components, Int Microwave
1696᎐11704. Symp Dig, 1998, pp. 9᎐12.
BIOGRAPHIES
Pietro Burrascano received his Laurea His research interests include: Artificial Neural Networks,
degree in electronic engineering in 1980 unsupervised learning theory, principal component analysis,
from the University of Rome ‘‘La independent component analysis, neural units with adaptive
Sapienza,’’ Italy, where he continued his activation functions; linear and non-linear adaptive discrete-
research activity until 1981. From 1981 to time filtering, blind deconvolution of non-minimum phase
1983 he was with Telespazio S.p.A., in- systems, blind image deblurring; non-destructive matter test
volved in planning activity of satellite and defect identification by neural networks based processing
telecommunication systems. In 1983 he of eddy current inspection date; vision and image processing
joined Infocom Dept. at the University of by neural networks for robot motion control.
Rome ‘‘La Sapienza’’ as a researcher and, after 1992, as an
Associated Professor of electrical circuit theory. From 1994 he
is Full Professor at the University of Perugia, Italy.
Mauro Mongiardo was born on Decem-
He is the author of several papers in the fields of techno-
ber 3, 1959 in Rome, Italy. He has stud-
logical development of thin film filters, digital circuits synthe-
ied at the University of Rome ‘‘La
sis, parametric modelling of physical systems, digital signal
Sapienza.’’ After obtaining the Laurea
processing and neural networks. His current research activi-
degree in 1983 he has been involved in
ties are concerned with theory and applications of artificial
joint projects between the University of
neural networks to signal processing.
Rome ‘‘Tor Vergata’’ and a local industry
ŽElettronica .. In 1988 he became an as-
Simone Fiori was born in Rimini in 1971. sistant professor in microwaves at the
He received the Italian Laurea with Hon- University of Rome ‘‘Tor Vergata’’; during that period he has
ors in Electronics Engineering in 1996, also studied for a Ph.D. degree at the University of Bath.
and is currently working toward his Ph.D. Since 1992 he has been an associate professor, first at the
in Electrical Engineering ŽCircuit The- University of Palermo and then at the University of Perugia.
ory. at the Department of Electronics His scientific interests have been mainly focused in the CAD
and Automatics of the University of An- and optimization of microwave and millimeter-wave passive
cona. In 1998 he joined the Department components and MMICs. He has served on the Technical
of Industrial Engineering of the Univer- Program Committee ŽCAD procedures. of the Microwave
sity of Perugia as Two-Year Visiting Researcher. Theory and Technique Symposium. He has also served as
174 Burrascano, Fiori, and Mongiardo
reviewer for several Journals and Conferences Ž IEEE MTT netism, AEU.. Mauro Mongiardo has been a visiting professor
Transactions, IEEE Microwa¨ e and Guided Wa¨ e Letters, Inter- at the Bath University, at the University of Victoria, B.C.,
national Journal of Numerical Modeling, International Journal Canada, the University of Corvallis, Oregon, USA, and the
of Microwa¨ e and Millimeter-Wa¨ e Computer-Aided Engineer- technical University of Muenchen, Germany. Mauro Mongia-
ing, Journal of Electromagnetic Wa¨ es, European Microwave rdo is the co-author of one book and about 140 technical
Conference, Italian National Conference of Electromag- publications.