Information on workshops (those available at the moment) are at the end of this document. You can find the appropriate abstract by clicking on the abstract title below.
Genetic Algorithms for Finance
Brian Aldershof, Lafayette College & Graham Capital Management
aldershb@lafvax.lafayette.edu
Genetic Programming (GP) is a widely known evolutionary technique that exploits genotype representations of variable size, namely programs. In addition, various versions of Genetic Algorithms (GA), Evolutionary Programming (EP), and Evolution Strategies (ES) also use variable length representations.
The variable size of representations makes analysis and understanding of evolutionary algorithms even more challenging than for the standard versions with fixed length genotypes. However, if variable size representations could be understood with respect to the evolutionary search process, the power of evolutionary algorithms would be greatly extended.
An intuitive advantage of variable length representations in evolutionary algorithms is that a larger, fitter genotype will evolve as a result of improvements being selected, promoted and combined. Ideally, a larger genotype should be due to selection of improvements instead of genetic operations with no effect. However, in practice this is not always the case. For example, in GP the size of evolved structures tends to grow indefinitely and is often not accompanied by improvements in fitness. Instead of the ideal scenario where evolution allows and favors the transmission of chunks of useful genetic information, a complicated process occurs: genetic operations tend to bloat the programs with unexpressed genetic material.
Why don't our algorithms work in the ideal manner? How can the evolution of a more complex solution be guided by modifying either the algorithm or representation? Can innovative practitioners offer specific representational choices that have proven useful when combined with specific algorithms?
More precisely this workshop invites discussions of:
Researchers already apply various techniques for controlling size by means of parsimony penalties to fitness. However, the penalty component changes the fitness landscape. Important questions arise about explicit or implicit, predefined or adaptive control of size growth, adaptativity of operators and representations. On this topic we expect contributions from participants who are jointly attending the intron workshop at ICGA on the following day.
Alternatively, a solution may lie in refining and more explicitly defining the genotype to phenotype mapping. Then the genetic operators which blindly manipulate genotypes may have greater odds of producing fitter, more modular and less bloated solutions.
We encourage the highlighting of tricks for controlling evolution in complex application domains. How well do such tricks work? How general are they? Are there theoretical or heuristic explanations why a particular trick works?
Please submit your paper electronically to:
VarSize-ICGA97@scr.siemens.com
Deadline for Submission: JUNE 8, 1997
Notification of Schedule: JULY 4, 1997
Workshop Date: SUNDAY, July 20: 7:30PM - 10:30PM
WORKSHOP FORMAT:
We plan an active participatory workshop. This will not be a series of short presentations of results with even shorter question periods!
Instead, we will open with a short overview presentation amalgamated from submitted papers that puts forth key issues and approaches to using variable length representations in evolutionary algorithms. This will be followed by a chaired discussion on various topics. The chair will start by soliciting participants whose submission is relevent to the topic and then open the floor to general discussion. When participants leave, our goal is that they have a better understanding of various avenues of approach as relayed by other participants' experience, and ideas for new projects.
Check this URL for updates: www.ai.mit.edu/people/unamay/icga-ws.html
We seek participation from researchers who are interested in the
emergence and effects of non-coding segments, the use of non-coding
segments in practical systems, and the relationship between biological
non-coding DNA and computational non-coding segments.
The workshop will consist of brief presentations of research followed
by an open discussion session. Researchers who are interested in
participating should send an extended abstract or position paper
(please limit to three pages) to one of the workshop organizers by
June 9, 1997.
Additional workshop information will be available at:
http://www.aic.nrl.navy.mil/~aswu/icga97.ws/
In essence, Evolutionary Programming models evolution at a more abstract
level than Genetic Algorithms. In particular, Evolutionary Programming
models evolution at the level of non-mixing species competing for a common
pool of limited resources. Because of the commitment to competing
non-mixing species, EP applications rarely if ever use recombination when
creating offspring but instead employ a wide variety of representation
specific mutations. This approach has been shown to provide significant
advantages for a variety of problems. EP has been shown to be effective as
a non-linear optimization technique and has a long history of evolving
complex behaviors using finite state machines, neural networks and most
recently programs represented as parse trees (i.e. crossoverless genetic
programming).
This workshop will give an overview of the state-of-the-art in
Evolutionary Programming and present a selection of topics currently under
investigation by researchers active in the EP field. Attendees are
encourgaed to participate through questions and discussions held after
each presentation.
Please check the URL http://www.natural-selection.com/eps/icga97.html
for more details.
The goal of this workshop is to provide many more test problem
generators to the EA community. To accomplish this, we need your
help! If you are interested in participating in or contributing to
this workshop, please send email either of the workshop organizers.
For more information, see
http://www.cs.gmu.edu/~mpotter/icga97-workshop.html.
We seek participation from researchers working in robotics and
evolutionary algorithms (GA, ES, GP, EP, etc). The format of the
workshop will be an overview of the field, followed by brief synopses of
current research, and then an open discussion session.
Topics of interest include (but are not limited to) the use of
evolutionary algorithms in:
The abstracts, and additional workshop information will be available at:
http://www.aic.nrl.navy.mil/~schultz/evolv-robots/
Workshop on Exploring Non-Coding Segments and Genetic-Based
Encodings
Researchers in the areas of genetic algorithms (GAs) and genetic
programming (GP) have recently begun to examine the utility of
non-coding segments in chromosome encodings. Non-coding segments,
which have also been called introns in GA/GP literature, refer to
regions on a chromosome that do not contribute to the overall fitness
of the chromosome. These structures are derived from non-coding DNA
which exists in abundance in natural systems. The fact that non-coding
DNA has not been selected out in spite of the extra energy required to
maintain it suggests that there must be some advantage to having it in
the genome. Research suggests that computational non-coding segments
may guard against the disruptive effects of crossover, provide natural
backups for the coding regions, and possibly expedite the evolution
of solutions.
Workshop on Evolutionary Programming
Evolutionary Programming was first defined in 1959 as a method for
inducing complex behaviors by evolving finite state machines. Since then,
Evolutionary Programming has expanded to include a number of related
topics including many of interest in the GA and GP communities. Although
EP and other evolutionary computations share many similarities, their
differences make each more suited for solving certain problems and
manipulating specific representations. The purpose of this workshop is to
explore what makes EP distinct as an evolutionary computation and
demonstrate how those differences provide advantages when solving certain
problems.
Workshop on Test Problem Generators for EAs
In comparing and contrasting the performance of evolutionary
algorithms (EAs), there is currently an over-reliance on individual
problems and/or rigid test suites of problems. An alternative
approach is to create test problem generators in which random problems
with certain characteristics can be generated automatically and
methodically. Some example characteristics would be multimodality,
epistasis, the degree of deception, and problem size. Since problems
are randomly created within a certain class, it is often easier to
draw general conclusions about the behavior of an EA. Furthermore,
the strengths and weaknesses of the algorithms can be tied to specific
problem characteristics.
Workshop on Evolutionary Robotics
Overlapping and complementary approaches to using evolutionary
algorithms in the areas of robotics and highly autonomous systems have
been seen in recent years. At the workshop, we would like to explore
some of these approaches, and hopefully better understand similarities
in and differences between these approaches.
Researchers interested in participating should send a description of
their research (please limit this to 300 to 500 words) to one of the
workshop organizers by 9 June.