NRLgate -
Plagiarism by Peer Reviewers
Sections 1.0 thru 1.8
This page is part of the NRLgate Web site presenting evidence of
plagiarism among scientific peer reviewers involving 9 different peer review
documents of 4 different journal and conference papers in the fields of
evolutionary computation and machine learning.
This page contains sections 1.0 through 1.8 of "Background Information."
Go to NRLgate Home Page
Go to Abbreviated Table of Contents
Go to Detailed Table of Contents
Go to Complaint Letter
to the Evolutionary Computation journal
Go to top of Next Page
1. Background Information
Scientific journals and conferences provide a mechanism for scientists to
report research results and exchange ideas in order to advance the progress
of science.
Publication in scientific journals and conferences is an important element
in the careers of most scientific researchers as reflected by the well-known
"publish or perish" principle. The presence or absence of publications
in scientific journals and published conference proceedings plays a pivotal
role in determining retention, promotion, and advancement in both the academic
and governmental research communities, in determining tenure in the academic
world, and in determining the funding of research projects in both the academic
and governmental research communities. Thus, a decision to accept a submitted
paper for publication translates directly into tangible benefits for the
submitting authors while a decision to reject is a tangible loss (particularly
in a journal or conference that is potentially viewed as being central to
the researcher's main field of work).
Most articles in most scientific journals and conference proceedings are
published after a public solicitation of papers and a competitive evaluation
process. The process of selecting articles for publication in scientific
journals is competitive (often highly competitive). Only a fraction (typically
between a quarter and a half) of submitted papers are accepted.
1.1. Genetic Programming (GP)
Genetic programming (GP) is a recently developed method for
automatically creating a computer program to solve a problem. Genetic programming
addresses one of the central questions of computer science --- namely, how
to get a computer to automatically solve a problem without having to explicitly
write, by hand, a detailed computer program.
By way of background, I am regarded as the developer and most prominent
proponent of the relatively new technique of genetic programming. My first
scientific paper on genetic programming appeared in 1989 at the International
Joint Conference on Artificial Intelligence (IJCAI-89). I am also author
of two books from the MIT Press on this subject, namely the 819-page book
Genetic Programming: On the Programming of Computers by Means of Natural
Selection published by the MIT Press (Koza 1992) and and the 746-page
book Genetic Programming II: Automatic Discovery of Reusable Programs
(Koza 1994a). A third book on genetic programming is in the works.
Genetic programming is a form of artificial intelligence and bridges two
fields:
- evolutionary computation and
- machine learning.
The term evolutionary computation (EC) refers to mathematical
problem-solving methods that are based on the Darwinian principle of survival
of the fittest and the ideas of evolution by means of natural selection.
Evolutionary computation encompasses the subfields of genetic algorithms
(GAs), genetic programming (GP), classifier systems, evolution strategies,
and evolutionary programming. Genetic programming is a variation of the
genetic algorithm. Over 5,000 scientific papers have been published in the
field of evolutionary computation (mostly in the last 15 years) and there
are over a thousand researchers around the world who are involved to some
extent in this field.
Machine learning (ML) is a subfield of artificial intelligence
(AI) that is concerned with automated methods by which computers learn to
solve problems.
Genetic programming is the fastest growing segment of the field of evolutionary
computation. Since 1989, there have been over 600 scientific papers published
in this field (including over 70 of mine). In 1996, the first conference
devoted to genetic programming was held with 292 registered attendees and
with 73 regular papers contained in a proceedings book published by The
MIT Press. Additional information about genetic programming and me can be
obtained at John
Koza's home page.
1.2. The Evolutionary Computation Journal (ECJ)
The Evolutionary Computation journal (ECJ) is currently the only
scientific journal devoted entirely to the field of evolutionary computation.
Since its inception in 1992 (and until just recently), the Evolutionary
Computation journal has had one editor-in-chief, three associate editors,
and 32 editorial board members.
Kenneth DeJong
has been Editor-In-Chief of the journal since the journal's inception in
1992 and for the entire time period involved herein. DeJong is employed
at Code (department) 5510 of the Naval Research Laboratory (NRL) in Washington,
D. C. He also teaches part-time in the Computer Science Department at George
Mason University in Fairfax, Virginia where he supervises several PhD students
in computer science (at least two of whom are full-time employees at Code
5514 and at least one of whom is a contractor of Code 5514).
John Grefenstette has
been North American Associate Editor of the journal since the journal's
inception. Grefenstette is Section Head of the Machine Learning research
staff at Code 5514 of the Naval Research Laboratory in Washington. His "biographical
sketch" at NRL states,
- As NRL's Machine Learning Section Head, Dr. Grefenstette manages
basic research programs in machine learning, establishes and maintains working
relationships with leading academic and corporate centers of AI research,
and develops applications of machine learning techniques to problems of
Navy interest. Dr. Grefenstette is responsible for the evaluation of available
machine learning techniques for application to problems of Navy interest,
and the identification of new areas of research that may be needed to bring
this technology to the application stage.
By way of background, John Grefenstette obtained his PhD from the University
of Pittsburgh in 1980 under the chairmanship of Dr. Kenneth DeJong who was
at the University of Pittsburgh at the time. In addition, Grefenstette is
an editor of the Machine Learning journal and Kenneth DeJong is a
member of its editorial board.
DeJong and Grefenstette serve (often in conjunction with one or more other
employees of NRL) as peer reviewers for various scientific conferences,
including the International Conference on Genetic Algorithms (ICGA), Foundations
of Genetic Algorithms workshop (FOGA), IEEE International Conference on
Evolutionary Computation (ICEC), Parallel Problem Solving from Nature conference
(PPSN), and the recently created International Conference on Evolutionary
Computation and its Applications (EvCA).
In addition, DeJong and Grefenstette serve (often in conjunction with one
or two other employees of NRL) as peer reviewers for the Machine Learning
Conference (next section).
Grefenstette is also on the editorial board of the Adaptive Behavior
journal.
The Evolutionary Computation journal is typical of scientific journals
and conferences in that its submitted papers go through a process of peer
review to decide whether they will be published. The peer reviewers of the
Evolutionary Computation journal are typically selected by the editors
from the journal's advertised editorial board. The 32 members of the journal's
editorial board are highly dispersed geographically and institutionally
(ten being from outside the United States). None of the 32 members are located
within the District of Columbia. The editorial board has never met as a
group, but, instead, communicates by electronic mail (e-mail) and ordinary
physical mail.
The relevant instructions to authors submitting papers to the Evolutionary
Computation journal call for the submitting author to send five physical
paper copies of his paper to the associate editor in the author's geographic
area. Thus, I submitted a paper applying genetic programming to electrical
circuit design (hereinafter referred to as the "ECJ paper") by
sending it by U. S. mail to North American Associate Editor John Grefenstette
at NRL.
For the first two years of the journal's operation, North American Associate
Editor Grefenstette handled about 80% of all papers submitted to the journal.
(The journal also had associate editors in Europe and Japan who handle papers
originating in their designated geographic areas, but they have no known
or likely connection with the subject matter of this document; accordingly,
the term "North American Associate Editor" herein refers to North
American Associate Editor John Grefenstette).
Upon receipt of a paper, the editors pick three peer reviewers and send
one physical copy of the paper to each reviewer (usually by U. S. mail).
At the same time, the journal sends the reviewer, by electronic mail, a
standard 7-part paper review form that asks the reviewers for his evaluation
of the paper. At the time of submission of my paper to the Evolutionary
Computation journal, it was generally understood that all peer reviews
were being done by members of the advertised editorial board; however, the
journal occasionally appoints peer reviewers from outside the editorial
board (and has recently done so with increasing frequency).
The peer reviewers normally read their physical paper copy of the submitted
paper and then write a review using the journal's 7-part paper review form.
The last two questions on the paper review form ask for the peer reviewer's
recommendation to the editors about whether the journal should (1) accept
the paper, (2) accept it on the condition that it be revised, or (3) reject
it. After the peer reviewers finish writing their reviews, they transmit
their review to the journal (typically by e-mail). The editorial leadership
of the journal then reads the three reviews written by the peer reviewers,
weighs the possibly conflicting recommendations from the three reviewers,
and reaches a decision on whether to publish the submitted paper.
The most common form of peer review for papers submitted to scientific conferences
and journals is "anonymous" in the sense that an author of a submitted
paper is ordinarily given the text of the reviews written by the peer reviewers
for his submitted paper, but not the names of the peer reviewers.
Out of 96 papers published by the Evolutionary Computation journal
during its first 3 full years, 3 have been on genetic programming.
1.3. The Machine Learning Conference (MLC)
The International Conference on Machine Learning is an annual scientific
conference that encompasses the field of machine learning.
The Machine Learning Conference (MLC) issues an annual Call For Papers (CFP)
with a paper submission deadline that is typically in January of each year.
Papers submitted to the Machine Learning Conference go through a process
of peer review to decide whether they will be published. Papers submitted
to the Machine Learning Conference are typically read by two or three scientific
peer reviewers drawn from the two dozen or so members of the conference's
program committee. The peer reviewers write a review of each paper using
the conference's standard paper review form. Finally, the program committee
receives the reviews written by the peer reviewers and decides which papers
will be accepted for presentation at the conference (typically held in June
or July of each year) and for publication in the proceedings book of the
conference (typically published at the time of the conference). The conference
meets in various locations (e.g., Italy in 1996, Tahoe in 1995).
The members of the program committee are more or less the same from year
to year. Most of the two dozen or so members of the program committee are
highly dispersed geographically and institutionally. There have been between
2 and 4 members of the MLC program committee from the Naval Research Laboratory
(NRL) since the late 1980s. For 1996, there were 4 members of the MLC program
committee from NRL. For almost every year (and, specifically, for the years
applicable to the particular MLC papers discussed herein), the members included
- Kenneth DeJong, PhD, of Code 5510 of NRL
- John Grefenstette, Ph.D., of Code 5514 of NRL
Out of approximately 25 submissions since 1990, only one paper (in 1994
by Rosca) involving the use of genetic programming has been published by
the Machine Learning Conference in its 13 year history. In 1995, all 7 submitted
papers involving the use of genetic programming were rejected. One paper
ridiculing genetic programming was published in 1995.
I have submitted papers to the Machine Learning Conference in various years
(two of which will be referred to herein as the "MLC papers").
The Machine Learning Conference (MLC) regularly receives grants each year
from the Office of Naval Research (ONR).
Many of the members of the program committee of the annual Machine Learning
Conference are also on the editorial board of the Machine Learning journal.
For example, John Grefenstette is an editor of the Machine Learning
journal and Kenneth DeJong is a member of its editorial board.
No paper on genetic programming has ever been published in the Machine
Learning journal.
1.4. Tools for Artificial Intelligence Conference (TAI)
The IEEE International Conference on Tools for Artificial Intelligence (TAI)
is an annual scientific conference. The subject matter of the TAI conference
is considerably broader than machine learning or evolutionary computation.
TAI includes a wide range of subjects from the field of artificial intelligence.
In the one year when I submitted a paper to the TAI conference (hereinafter
referred to as the "TAI paper"), papers on evolutionary computation
papers represented only a tiny fraction (8 out of 120) of the published
papers. In that year, there were only 4 reviewers on TAI's "List of
Reviewers" who were involved in any known way in the field of evolutionary
computation. All 4 of the EC-knowledgable reviewers at the TAI conference
had the same institutional affiliation, namely the Naval Research Laboratory:
- Kenneth DeJong, of Code 5510 of NRL
- John Grefenstette, of Code 5514 of NRL
- Alan C. Schultz, of Code 5514 of NRL
- William M. Spears, of Code 5514 of NRL
1.5. Other conferences in evolutionary computation
The International Conference on Genetic Algorithms (ICGA) is a biannual
conference covering all aspects of evolutionary computation, including genetic
algorithms, genetic programming, classifier systems, evolution strategies,
and evolutionary programming.
The conference committee for the first International Conference on Genetic
Algorithms (ICGA-85) consisted of John Grefenstette (then at Vanderbilt
University), Kenneth DeJong of NRL, Lashon Booker of NRL, Stephen F. Smith
of CMU in Pittsburgh and Prof. John Holland of the University of Michigan
(the inventor of both the genetic algorithm and the genetic classifier system).
By way of background, Lashon Booker is a former employee of Code 5510 of
the Naval Research Laboratory and is currently working for a military contractor
in suburban Virginia.
David Schaffer has co-authored papers with Grefenstette during the period
when both were at Vanderbilt University in the mid 1980's, and now works
in upstate New York at Philips Laboratories.
Stephen F. Smith (who also received his PhD at the University of Pittsburgh
under Kenneth DeJong in 1980 and has been at Carnegie Mellon University
(CMU) in Pittsburgh since the early 1980's).
In 1987, the ICGA-87 program chair was John Grefenstette (then at NRL) and
the program committee included Kenneth DeJong (NRL), Lashon Booker (NRL),
and Stephen F. Smith of CMU in Pittsburgh and four others. In 1989, Kenneth
DeJong was the conference chair; David Schaffer was the program chair; John
Grefenstette was finance chair; and the conference committee included Lashon
Booker. In 1991, Kenneth DeJong and David Schaffer were conference co-chairs,
and Lashon Booker was one of two program co-chairs.
The Navy Research Laboratory has made financial contributions to each ICGA
conference since 1985. The Office of Naval Research (ONR) has made financial
contributions to ICGA in recent years.
In addition, Grefenstette, DeJong, and other NRL personnel are regularly
on the program committees and are reviewers for numerous other conferences
and workshops in the fields of evolutionary computation and machine learning,
including
- IEEE International Conference on Evolutionary Computation (ICEC) whose
program committee included Kenneth DeJong and William M. Spears, M.S., an
employee of Code 5514 (as well as David Schaffer and another employee of
Philips Laboratories) for its most recent annual conference in Nagoya,
- Evolutionary Programming conference (EP) whose program committee included
Kenneth DeJong and William M. Spears, M.S., an employee of Code 5514
- Parallel Problem Solving from Nature conference (PPSN) whose steering
committee included Kenneth DeJong and William M. Spears, M.S., an employee
of Code 5514 at its 1994 biannual conference,
- Foundations of Genetic Algorithms workshop (FOGA) whose program committee
included Kenneth DeJong and Grefenstette at its 1994 biannual meeting,
- International Conference on Evolutionary Computation and its Applications
(EvCA) whose program committee included Kenneth DeJong at its first conference
in 1996.
1.6. The Naval Research Laboratory (NRL)
The Naval Research Laboratory (NRL) is a well-regarded institution that
does research and development work within the Department of Navy. Its commanding
officer is Captain Bruce
W. Buckley, USN. Dr. Timothy
Coffey is the Director of Research at NRL.
Code 5510 is involved with research into artificial intelligence and is
also known as the Navy
Center for Applied Research in Artificial Intelligence (NCARAI). Dr.
Alan L. Meyrowitz has
been the director of Code 5510 since 1991. Kenneth DeJong is Chief Scientist
in Code 5510.
Code 5514 is one of four research groups within Code 5510. Code 5514 specializes
in machine learning. John Grefenstette is the Section Head of Code 5514.
The Office of Naval Research (ONR) funds a substantial part of the activity
at the Naval Research Laboratory in Washington, DC by contracting with NRL
to perform work. NRL is an internal research resource within the Department
of the Navy. NRL is consistently funded by ONR from year to year. In a loose
sense, ONR is the customer and NRL obtains funds by selling ONR work on
particular research and development projects.
1.7. The SAMUEL system developed in-house at the Naval Research Laboratory
(NRL)
One of the missions of the Naval Research Laboratory is to seek out emerging
new scientific technologies and to study them for possible usefulness to
the Navy.
For example, the NRL "biographical sketch" of John Grefenstette
of Code 5514 of the Naval Research Laboratory states,
- Dr. Grefenstette is responsible for the evaluation of available machine
learning techniques for application to problems of Navy interest, and the
identification of new areas of research that may be needed to bring this
technology to the application stage.
In some instances, staff members of codes 5510 and 5514 go beyond "evaluation
of available machine learning techniques" and have invented new technologies
in-house at NRL.
For example, John Grefenstette of Code 5514 of NRL is the inventor and the
most prominent proponent of a system for machine learning called SAMUEL
(Grefenstette 1989, 1991, 1992). SAMUEL addresses some of the same problems
that are addressed by genetic programming (of which I am regarded as the
most prominent proponent). For example, SAMUEL is particularly suitable
for developing control strategies for controlling autonomous robots and
for developing strategies for pursuer-evader games (evasive maneuvers).
Code 5514 of the Naval Research Laboratory has expended considerable effort
on the SAMUEL system since the late 1980s. On several hundred occasions,
NRL personnel have traveled extensively to give talks about SAMUEL as part
of a "technology transfer" effort from government laboratories
to university and other scientific audiences. Numerous scientific papers
have been published on the SAMUEL system at various conferences and journals
covering the fields of evolutionary computation and machine learning. The
authors of the papers on SAMUEL are almost always affiliated with NRL and
include John Grefenstette, Kenneth DeJong, various NRL employees (e.g.,
Connie Ramsey and Alan Schultz of Code 5514), and various NRL contractors
(e.g., Mitchell Potter of Code 5514) (Potter, DeJong, and Grefenstette,
1995; Schultz and Grefenstette 1990; Grefenstette, Ramsey, and Schultz 1990).
1.8. The connection between the scientific peer review process and the
problem of measuring the internal performance at the Naval Research Laboratory
(NRL)
High-level policy makers who manage the work of scientists and engineers
working with numerous different technologies face a difficult problem in
objectively evaluating such work.
The federal government spends a considerable amount of money on research
and development in machine learning and evolutionary computation. However,
nearly all of its is allocated by the government to agencies within the
government (such as Code 5514).
In allocating funds to its various departments and in making decisions about
the retention and advancement of scientific personnel, the high-level management
of both Naval Research Laboratory and ONR has developed numerous ways to
try to objectively measure internal performance.
Information from the outside scientific community has the obvious disadvantage
of originating from sources that are not directly attuned to the mission
of the Naval Research Laboratory; however, such outside information has
the alluring advantage to the Naval Research Laboratory management of appearing
to be arms-length and objective. Of course, the usefulness of the measurements
provided by the peer review process by the outside scientific community
depends on the extent to which these measurements are not affected by the
actions of the very departments and personnel whose work is being measured.
Publication by scientific conferences and journals of the research papers
written by Naval Research Laboratory authors is one measure of scientific
merit. The number of citations to papers is another measure of scientific
merit (that carries significant weight at Naval Research Laboratory apparently
because it is thought that citations are an indicator of successful "technology
transfer").
However, the affirmative appearance of publications by NRL authors is not
the most important part of the publication process.
The absence of published papers in the scientific literature that address
problems of interest to Naval Research Laboratory strongly supports in-house
funding in such seemingly neglected problems.
Thus, the funding of a department may depend on the extent to which a department
can create the impression to its own management that non-governmental scientists
are neglecting a particular field.
It should be noted that the SAMUEL technique (previous
section) that was invented in-house in Code 5514 of the Naval Research
Laboratory and the technique of genetic programming (GP) are rival technologies
within the fields of machine learning and evolutionary computation that
were each developed in the late l980s to solve artificial intelligence problems.
Author: John R. Koza
E-Mail: NRLgate@cris.com
Go to NRLgate Home Page
Go to Abbreviated Table of Contents
Go to Detailed Table of Contents
Go to Complaint Letter
to the Evolutionary Computation journal
Go to top of Next Page