IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 16, NO.

4, NOVEMBER 2001 757

Power System Security Assessment Using

Neural Networks: Feature Selection Using Fisher
Discrimination
Craig A. Jensen, Mohamed A. El-Sharkawi, and Robert J. Marks, II

Abstract—One of the most important considerations in ap- states that, as a rule of thumb, the required cardinality of the
plying neural networks to power system security assessment is training set for accurate training increases exponentially with
the proper selection of training features. Modern interconnected the input dimension [5]. Thus, choosing a small subset of the
power systems often consist of thousands of pieces of equipment
each of which may have an affect on the security of the system. thousands of possible features, i.e. feature selection, requires
Neural networks have shown great promise for their ability to a small fraction of the training samples required if all features
quickly and accurately predict the system security when trained are used. Feature selection is the process of identifying those
with data collected from a small subset of system variables. This features that contribute most to the discrimination ability of the
paper investigates the use of Fisher’s linear discriminant function, neural network. Only these features are then used to train the
coupled with feature selection techniques as a means for selecting
neural network training features for power system security neural network and the rest are discarded. Proposed methods
assessment. A case study is performed on the IEEE 50-generator for selecting an appropriate subset of features are numerous
system to illustrate the effectiveness of the proposed techniques. [1]–[4].
Index Terms—Dynamic security, intelligent systems, neural net- An alternate to feature selection is feature extraction. Here,
work, power system. the dimensionality of a feature set is reduced by combining fea-
tures while retaining characteristics that allow for accurate clas-
sification. Feature extraction is the process of mapping all avail-
I. INTRODUCTION
able features into a composite feature set of lower dimension.

T HE trend toward deregulation has forced modern utilities

to operate their systems closer to security boundaries. This
has fueled the need for faster and more accurate methods of se-
Feature extraction techniques applied to power system security
assessment have previously been presented by Weerasooriya
and El- Sharkawi [6].
curity assessment. Power system security assessment deals with Many feature extraction techniques such as the principle com-
the system’s ability to continue to provide service in the event of ponents algorithm are based on the assumption that the greater
an unforeseen contingency. Such contingencies may include the the spread of the data in a particular axis, the greater the ef-
unexpected loss of an important transmission circuit or a sudden fect that will have on the discrimination ability of the neural
change in a large load. Either of which could lead to a disrup- network. This need not be true. Feature selection methods, on
tion of service to part or all of the system. The goal of security the other hand, generally are based on ranking different combi-
assessment is to determine when a disruption of service is likely nations of features in accordance to their classification perfor-
to occur and to take steps to reduce the risk. mance and choosing the combination that achieves the highest
Neural networks have shown great promise as a means of ranking. Unlike feature extraction, no preprocessing is required
predicting the security of large electric power systems [1]–[3]. once the features are chosen.
Neural networks offer several advantages over traditional tech- The basic problem encountered in all feature selection algo-
niques including the ability to learn from examples. rithms is how to deal with the computational complexity in-
The first step in applying neural networks to power system volved in searching among the large number of possible solu-
security assessment is the creation of an appropriate training tions. Given a set of features , , the task is to
data set. A common approach is to simulate the system in determine the subset of features that best satisfies some
response to various disturbances and then collect a set of selection criteria. The search for the optimal subset of features is
pre-disturbance system features along with the corresponding a combinatorial problem that requires selection among pos-
system security index. Possible security indices include the sible solutions, where,
Critical Clearing Time (CCT) [1] and the system Energy Margin
(EM) [3]. One of the most important aspects of achieving good
neural network performance has proven to be the proper
selection of training features. The curse of dimensionality becomes excessive even for relatively small values of and
.
Manuscript received November 29, 1999; revised June 12, 2001. This work The simplest method to deal with this complexity, involves
was supported by the National Science Foundation. determining a goodness score for each individual feature, and
The authors are with the Department of Electrical Engineering, University of
Washington, Seattle, WA 98195. then ranking the features based on their scores. A commonly
Publisher Item Identifier S 0885-8950(01)09421-4. used example of a goodness score uses the Fisher discriminant.
0885–8950/01$10.00 © 2001 IEEE
758 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 16, NO. 4, NOVEMBER 2001

The between-class scatter, , is

Solving (1) for the Fisher weights, , that maximize can

then be written as

Fisher’s linear discriminant function is a projection from a

Fig. 1. (a) The Fisher linear discriminant projects data onto a lower dimension -dimensional space onto a line in such a manner that the
that maximally separates the classes. (b) A nonoptimal discriminant function. training data is best separated. Fig. 1(a) shows the optimal
(Fisher) projection for a 2-dimensional problem and Fig. 1(b)
The simple ranking method gives poor performance in cases shows a nonoptimal projection. It can be seen that the Fisher
when many features are highly correlated. In such cases, sev- vector, , defines the line — or, more generally, the hyperplane
eral features may have high individual discrimination ability, but – that results in the greatest separation between the ’s and ’s
due to their correlation, they offer no improvement when used when data are linearly.
together. For this reason, more advanced methods are presented The Fisher discriminant can also be used as a means of as-
to find features that work well together. sessing the linear separability of two classes of data. For ex-
Section III presents the IEEE 50-genrator transient stability ample, the magnitude of the Fisher function, , can be used as
test system. Section IV outlines several feature selection tech- a measure of how well a linear classifier will perform on a given
niques that have been proposed. In Section V, a case study of the problem. The higher the value of , the more separable the data
proposed techniques is presented to illustrate how neural net- are, and thus the higher the expected performance of the clas-
work training features can be selected that are independent of sifier. Different combinations of features can also be compared
changes of system topology. Finally, conclusions are presented by comparing their respective Fisher values. This technique is
in Section VI. used in the proceeding sections as a method for comparing dif-
ferent combinations of features in an effort to find the optimal
II. FISHER’S LINEAR DISCRIMINANT combination.
The Fisher linear discriminant function was proposed in 1936
[6],[7]. It seeks to find the optimal linear discriminant func- III. FEATURE SELECTION TECHNIQUES
tion for separating two classes of data. Given a set of - If the dimension of the problem is small, a reasonable tech-
dimensional training samples with samples in nique is to simply search through all possible combinations of
class and samples in class , the task is to find the linear features and calculate their corresponding Fisher values. The
mapping, , that maximizes, optimal feature set is then the set with the highest Fisher value.
This is only practical on small problems as it requires the eval-
uation of Fisher values.
In problems where the dimension is too high for an exhaustive
where is the mean of class and is the variance of . In search, a search technique called the backtrack search [4], also
words, this corresponds to finding the line that when the data are known as the branch-and-bound algorithm, can be employed to
projected onto, provides the maximum separation (see Fig. 1). ease the computational burden. The backtrack search guarantees
The criterion function, , can be rewritten as an explicit func- the optimal solution if the criterion function satisfies the mono-
tion of as tonicity condition. The monotonicity condition requires that the
value of the criterion function be nondecreasing when additional
(1) features are added. The Fisher linear discriminant function sat-
isfies this condition because increasing the input dimension, ,
where is referred to as the between-class scatter matrix and can never result in a decrease in the Fisher value .
is the within-class scatter matrix. The within-class scatter, The mechanics of the backtrack search are best illustrated, as
, is defined as in Fig. 2, in the form of a search tree. The root of the tree con-
tains all of the features to be considered. Each node under the
root corresponds to the removal of a single feature. The tree has
where and are depth , and the leaves of the tree contain all possible com-
binations of features out of the original . The algorithm is
initialized by calculating the Fisher value at any one of the leaves
and the value is stored as the current best solution. The tree is
and traversed starting at the root and calculating the Fisher value,
, at each node along the way. Since the monotonicity condi-
tion applies, the value of is guaranteed to be nonincreasing as
the tree is traversed Therefore, if at any point along the traversal,
JENSEN et al.: POWER SYSTEM SECURITY ASSESSMENT USING NEURAL NETWORKS: FEATURE SELECTION USING FISHER DISCRIMINATION 759

There are several variations including the forward sequential,

backward sequential and the plus-l take away-r algorithm.
The forward sequential method is a bottom-up algorithm that
starts by choosing the best individual feature. Then the feature
set is built from the ground up, by repeatedly adding the next
feature that works best with the previously chosen features. This
algorithm requires evaluations. The backward
sequential method is very similar to the forward sequential ex-
cept that it is a top-down algorithm. Instead of starting with an
empty set and building a set of features, the backward sequen-
tial method starts with the complete set of features and itera-
tively removes them one at a time until only the desired number
Fig. 2. Search tree for the backtrack algorithm for choosing the 2 best features remain. This is done by calculating the Fisher value for all sub-
out of 5.
sets of features where is the current number of features
remaining. The feature that results in the smallest decrease in
the value of falls below current best value found at any of the Fisher value at each iteration is then removed. The backward
leaves, the rest of that branch need not be evaluated. sequential method requires the same number of evaluations as
An example illustrating the backtrack search when applied the forward method, but is more computationally demanding be-
to the selection of the two best features out of five is shown in cause the calculation of the Fisher discriminant is in a higher
Fig. 2. Initially, the branch at the far right side is evaluated and dimensional space at each iteration.
the Fisher value of the subset is determined and saved It should be noted that while the Fisher linear discriminant
in the variable . Next, the algorithm backtracks and proceeds function has been shown to produce the optimal classification
down the next unexplored branch comparing the Fisher value, boundary for a linear classifier, its performance on a nonlinear
, at each node to . If the value of exceeds , there is still neural network classifier is not known. The nonlinear neural
a chance that a better feature set will be discovered, so the al- network is commonly accepted as a more general model than
gorithm must continue to search along the rightmost unexplored a linear classifier. The plus- take away- algorithm is essen-
branch. When the bottom of the tree is reached, if the value of tially a combination of the forward sequential and backward
is still greater than , then is updated accordingly. If on the sequential methods. In this algorithm, features are added via
other hand, a node is found where the value of is less than , the forward method followed by the removal of features by
the rest of that branch need not be traversed due to the mono- the backward method. This is advantageous in that it provides
tonicity condition. After the entire tree has been evaluated the the algorithm the ability to remove features that are no longer
optimal feature set is the one corresponding to . needed due to combinations of features subsequently added. The
The backtrack method is particularly efficient when the fea- algorithm works in either the top-down or bottom-up direction,
tures are preordered and placed in the tree in order of the mag- depending on if or . For the case where, , the
nitude of their individual Fisher values from left to right. This order of the forward-backward methods is be reversed, i.e., the
tends to put the better solutions toward the right side of the tree, algorithm starts with all the features and features are removed
which, in turn, tends to allow the algorithm to eliminate the max- followed by the addition of features.
imum number of branches from the search.
The backtrack method should be used with caution because
IV. IEEE 50 GENERATOR SYSTEM
its worst-case run time can be much worse than exhaustive
search. In worst case, the backtrack algorithm must evaluate To assess the applicability of the Fisher linear discriminant
every node in the tree. The number of evaluations required in function as a means of selecting neural network training
the worst case is therefore, features, an empirical study was conducted on the IEEE 50-
generator transient stability test system [13]. The system con-
sists of 50 generators, 145 buses and 453 transmission lines and
transformers. A training database was collected by simulating
3-phase faults at nine different buses throughout the system.
Of the nine faults, 4 were on the 500 KV subsystem, 1 was
where on the 220 KV system and 4 were on the 100 KV subsystem.
total number of raw features; The locations were chosen to cover a wide cross-section of the
system. The critical clearing time for each fault was determined
number of features to be selected;
by simulating the system response using the EPRI Electric
height of the nodes in this level of the tree. Transient Mid-Term Stability Program [12] (ETMSP) software
In cases where the backtrack method proves computation- program. The system was simulated for 651 different system
ally infeasible, a sub-optimal solution needs to suffice. Sequen- operating conditions. The real and reactive power output of
tial ranking methods are relatively simple methods utilizing a each of the 50 generators as well as the total system real and
heuristic that involves building a feature set by choosing the reactive load level were recorded along with the system critical
next feature that works best with the previously chosen features. clearing time for each of the 651 simulations.
760 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 16, NO. 4, NOVEMBER 2001

TABLE I
STATISTICAL ANALYSIS OF FISHER’S LINEAR DISCRIMINANT APPLIED TO FEATURE SELECTION FOR SECURITY ASSESSMENT.

Fig. 3. Histograms showing the distribution of the neural network training error for various combinations of 4 randomly chosen features.

V. JUSTIFICATION OF FISHER FEATURE SELECTION for a linear classifier? This question is not easy to answer the-
oretically since, short of exhaustive search, there is no known
The Fisher forward sequential method was used to select the technique for determining the optimal feature set for a multi-
best 4 features for each of the nine faults considered in the study. layer perceptron neural network. What can be done, however,
Then a neural network was trained for each of the faults using is to perform a series of computer simulations to assess the dis-
the 4 best features found from the forward sequential method. crimination ability of a given feature set on a given neural net-
The neural network training results are summarized in Table I. A work model. These simulations can then be used to compare dif-
commercial neural network simulation package called QwikNet ferent feature sets and a statistical framework can be developed
[11] was used to train and test the neural networks. to quantify the performance of the Fisher linear discriminant.
Therefore, its performance should be at least as good as that Many statistical techniques exist for defining confidence in-
of the linear classifier. The question then arises as to how well tervals based on known probability distribution functions [9].
the neural network performs given the optimal set of features In this case, the probability distribution function of the neural
JENSEN et al.: POWER SYSTEM SECURITY ASSESSMENT USING NEURAL NETWORKS: FEATURE SELECTION USING FISHER DISCRIMINATION 761

network classification error for various combinations of input TABLE II

features can be estimated by randomly generating different fea- COMPARISON OF THE PROBABILITY OF THE EXISTENCE OF A FEATURE
SET BETTER THAN THE SET SELECTED BY THE FISHER FEATURE
ture combinations and determining their corresponding neural SELECTION ALGORITHM
network classification error. The neural network model used in
this study consisted of 3 layers with 4 inputs, 10 hidden neurons
and a single output. For each of the nine faults, 1000 random
combinations of 4 features were generated. A neural network
was then trained using each of the 1000 feature combinations.
Fig. 3 is a histogram of the neural network training error based
on randomly selecting 4 power system input features out of a
total of 102. The training error for each network was based on
the average training error over 5 training sessions with different
random initial weights. The results are summarized numerically
in Table I.
Several techniques exist for calculating tail probabilities of
the form , where is an unknown random variable
and is a given distance from the mean. If the variance, , and
the mean, , of the underlying process are known, the Cheby-
shev inequality [9] can be used to establish bounds on the prob-
ability of a sample occurring greater than a given distance, ,
from the mean. The Chebyshev inequality states that,
TABLE III
TOPOLOGY CHANGES

The Chebyshev inequality makes no assumptions regarding

the distribution of the underlying process, other than its mean
and variance. Because of this, the established bounds are often
very loose. Nevertheless, the Chebyshev inequality is still useful
for establishing a theoretical upper bound on the probability that
a feature combination exists that is superior to the combination
determined by the Fisher feature selection method. The right
column of Table I shows the probability of the existence of a
combination of features superior to that found by the Fisher for-
ward sequential technique.
Confidence intervals [10] are a well-known metric for as-
sessing the reliability of experimental results. The equations for
the upper and lower bounds on experimental results are,

(2)

where
point estimate of the trial;
number of samples;
of the normal distribution. Fig. 4. A portion of the high voltage transmission system surrounding the fault.
(2) can be used to place a bound on the observed results with The fault is on bus #7 and cleared by removing the line between buses 6 and 7.
Dashed lines indicate several of the transmission lines that were removed to
percent certainty. Experimental results show that the features simulate topology changes.
selected by the Fisher technique were found to be better than all
but 46 out of 9550, or 99.54% of all randomly generated feature
VI. CASE STUDY – APPLICATION TO TOPOLOGY CHANGES
combinations. The 95% confidence interval is (99.38, 99.65).
Table II shows the experimental 95% confidence intervals for In this section, a case study is performed on the IEEE 50-
each of the faults under study. generator system to show how the Fisher feature selection
762 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 16, NO. 4, NOVEMBER 2001

TABLE IV
COMPARISON OFNEURAL NETWORK PERFORMANCE BASED ON FEATURES SELECTED FROM A
SINGLE SYSTEM TOPOLOGY AND MULTIPLE SYSTEM TOPOLOGIES

technique can be used to select features that are independent of topology changes were spread across a wide area of the system
system topology. The system topology refers to the operating in an attempt to determine the significance of the location of
status of the numerous devices connected to the system. the topology change with respect to the fault. Some topology
Selecting features that are independent of topology is important changes consisted of removing 500 KV lines close to the fault
since the topology of modern power systems is continuously while others were 100 KV lines far removed from the fault.
changing. These changes are due to many factors including Table III shows the location and voltage level of each of the 11
maintenance, repair, and the addition of new equipment. topology changes.
Since the system topology is not fixed, special care needs be Fig. 4 shows a one-line diagram of a portion of the high
taken to minimize the effect of topology changes on the perfor- voltage transmission system surrounding the fault at bus #7.
mance of the neural network. One approach is to train different Several of the transmission lines that were removed as a result of
neural networks for each change in topology, and then use the the topology changes are shown as dashed lines, the rest occur
specific neural network that reflects the current topology of the further from the fault and are not shown.
system. This approach is only applicable to problems that in- The training data files for each neural network were created
volve relatively few changes in topology. If a large number of from the same raw data file, which included a fixed pre-fault
topology changes need to be considered, the number of neural system topology. The first neural network was trained with fea-
networks required becomes very large, and the problem quickly tures 112-Q, 104-P and 111-P, while the second neural network
becomes impractical. used features 111-Q, 104-P and 110-Q. The features correspond
Another approach is to choose features for training the neural to the either the real (P) or reactive (Q) generator power outputs
network that are independent of changes in topology. This al- at the given buses. As previously mentioned, the features for
lows a single neural network to learn the security of the system the first neural network were selected by the Fisher method ap-
with respect to various different system topologies. Features plied to a data file consisting of a fixed system topology, while
such as the aggregate generation in a specific area or the flow the features for the second neural network were derived from a
on important transmission circuits have been suggested [3]. The data file with multiple system topologies. Both neural networks
Fisher feature selection technique can be used to select features were tested with a data file composed of 641 patterns with 11
that reflect changes in topology, and therefore, provide good different system topologies. Five training runs were performed
performance even in the event of unexpected topology changes. for each network to assess their performance on the testing file.
This can be done by applying the Fisher selection technique The results are shown in Table IV.
to a training database that contains examples of as many dif- It can be seen from Table IV that the neural network trained
ferent system topologies as possible. The Fisher selection crite- with features selected from data containing multiple system
rion then selects the features that work best with regard to the topologies resulted in nearly a full order of magnitude reduction
various topologies. in testing error. This is a remarkable result considering the
An experiment was performed to test the performance of the neural network was trained with data consisting of only a single
Fisher technique for selecting topology invariant features for system topology! This test clearly shows the importance of
neural network training. The experiment consisted of training selecting features that are independent of changes in system
two neural networks with features selected by the Fisher topology and shows that the Fisher selection technique can be
forward sequential technique. The networks were trained with effectively used to select such features.
data generated from a single 3-phase fault located at bus #7
and cleared by removing the line between buses 7 and 6. The
VII. CONCLUSION
features for the first neural network were selected based on
a training database generated with a fixed system topology. The Fisher linear discriminant function coupled with the
The features for the second neural network were selected from sequential feature selection technique has been proposed as a
a training database containing 11 different pre-fault system means for selecting neural network training features for power
topologies. The topology changes involved removing various system security assessment. Through a statistical analysis, it
transmission circuits from the system and then simulating was shown that the Fisher technique was able to select a feature
the fault at bus #7 and calculating the system security. The set in the top one percentile of all possible feature sets. A case
JENSEN et al.: POWER SYSTEM SECURITY ASSESSMENT USING NEURAL NETWORKS: FEATURE SELECTION USING FISHER DISCRIMINATION 763

study was performed to illustrate how the Fisher technique Mohamed A. El-Sharkawi is a Fellow of IEEE. He received the B.Sc. degree in
can be used to select neural network training features that are Electrical Engineering in 1971 from Cairo High Institute of Technology, Egypt.
He received the M.A.Sc and Ph.D. degrees in Electrical Engineering were from
independent of changes in power system topology. The pro- the University of British Columbia in 1977 and 1980, respectively. In 1980,
posed methods were tested on the IEEE 50-generator transient he joined the University of Washington and is presently a Professor of Elec-
stability test system and excellent results were demonstrated. trical Engineering. Dr. El-Sharkawi is the founder of the international confer-
ence on the Application of Neural Networks to Power Systems (ANNPS), and
a co-founder of the international conference on Intelligent Systems Applica-
REFERENCES tions to Power (ISAP). He is a member of the administrative committee of the
IEEE Neural Networks Council representing the Power Engineering Society,
[1] D. J. Sobajic and Y. H. Pao, “Artificial neural-net based dynamic security and the multi-media Tutorial Chair of the IEEE Neural Networks Council. He
assessment for electric power systems,” IEEE Trans. Power Systems,
is a founding Chairman of several IEEE task forces and working groups and
vol. 4, no. 1, pp. 220–228, Feb. 1989. subcommittees, including the task force on Application of Neural Networks to
[2] M. A. El-Sharkawi, R. J. Marks II, M. E. Aggoune, D. C. Park, M. J.
Power Systems, the working group on Advanced Control Strategies for dc-type
Damborg, and L. E. Atlas, “Dynamic security assessment of power sys- Machines, and the task force on Intelligent Systems Application to Dynamic
tems using back error propagation artificial neural networks,” in Second Security Assessment. He is a co-founder of the IEEE Subcommittee on Intelli-
Symposium on Expert System Application to Power Systems, Seattle,
gent Systems. He is a current or past member of the editorial board or associate
WA, 1989. editor of several journals including the IEEE Transactions on Neural Networks,
[3] Y. Mansour, E. Vaahedi, and M. A. El-Sharkawi, “Dynamic security
the Engineering Intelligent Systems, and the International Journal of Neuro-
contingency screening and ranking using neural networks,” IEEE Trans. computing. Dr. El-Sharkawi is the co-editor of the IEEE tutorial book on the
Neural Networks, vol. 8, no. 1, pp. 942–950, July 1997.
applications of NN to power systems, and the author of a textbook on “funda-
[4] J. Kittler, A. Etemadi, and N. Choakjarernwanit, “Feature selection and mentals of Electric Drives” published by Brooks/Cole in 2000. Dr. El-Sharkawi
extraction in pattern recognition,” in Pattern recognition and Image pro- was the chairman of the IEEE International Electric Machines and Drives held
cessing in physics, R. A Vaughan, Ed., 1990, Proceedings of the 37th in Seattle, May 1999.
Scottish University summer school in physics. He organized and taught several international tutorials on intelligent systems
[5] R. D. Reed and R. J. Marks II, Neural Smithing: Supervised Learning in
applications, power quality and power systems, and organized and chaired nu-
Feedforward Artificial Neural Networks. Cambridge, MA: MIT Press, merous panel and special sessions in IEEE and other international conferences.
1999..
He has published over 140 papers and book chapters in these areas and holds
[6] S. Weerasooriya and M. A. El-Sharkawi, “Use of Karhunen-Loe’ve ex- 5 licensed patents:
pansion in training neural networks for static security assessment,” in
Proceedings of the First International Forum on Applications of Neural
Networks to Power Systems, Seattle, WA, July 1991, pp. 59–64.
[7] R. A. Fisher, “The use of multiple measurements in taxonomic prob-
lems,” Annals of eugenics, vol. 7, pp. 179–188, 1936.
[8] R. O. Duda and P. E. Hart, Pattern Classification and Scene Analysis:
John Wiley and Sons, 1973.
[9] A. Leon-Garcia, Probability and Random Processes for Electrical En-
gineering, 2nd ed: Addison-Wesley, 1994.
[10] H. J. Larson, Introduction to Probability Theory and Statistical Infer-
ence: John Wiley and Sons, 1969.
[11] C. A. Jensen. (1999) Qwiknet Neural Network Design Software Version
2.22 [Online]. Available: http://www.kagi.com/cjensen
[12] Extended Transient Midterm Stability Program, Version 3.0, Palo Alto,
California, vol. 1–6, 1993.
[13] “Transient stability test systems for direct stability methods,” IEEE
Trans. Power Systems, vol. 7, no. 1, pp. 37–44, Feb. 1992. Robert J. Marks, II is a Professor and Graduate Program Coordinator in the
Department of Electrical Engineering in the College of Engineering at the Uni-
versity of Washington, Seattle. He is Fellow of both IEEE and The Optical So-
ciety of America.
Craig A. Jensen received the BSEE and MSEE degrees from the University Dr. Marks served as the first President of the IEEE Neural Networks Council.
of North Dakota in 1993 and 1995, respectively, and the Ph.D. degree from the In 1992, he was given the honorary title of Charter President. He served as the
University of Washington in 1999. He is currently working for Microsoft Com- Editor-in-Chief of the IEEE TRANSACTIONS ON NEURAL NETWORKS
pany as a software engineer. He was employed at Otter Tail Power Company in and as a topical editor for Optical Signal Processing and Image Science for the
Fergus Falls, MN as an Engineering intern in the System Planning and System Journal of the Optical Society on America A. Space limitations and a keen sense
Engineering departments. His research interests are neural networks, computa- of modesty prohibit listing the rest of his remarkable professional achievements.
tional intelligence, intelligent systems applications to power systems and com- Dr. Marks is the co-author of the book entitled “Neural Smithing: Supervised
puter programming. Learning in Feedforward Artificial Neural Networks” (MIT Press, 1999)