Content uploaded by Lee Spector
Author content
All content in this area was uploaded by Lee Spector on Oct 10, 2017
Content may be subject to copyright.
Recent Developments in Autoconstructive Evolution
Lee Spector
School of Cognitive Science
Hampshire College
893 West Street
Amherst, Massachuses 01002
lspector@hampshire.edu
Eva Moscovici
College of Information and Computer Sciences
University of Massachuses Amherst
140 Governors Drive
Amherst, MA 01003
evamoshkovich92@gmail.com
ABSTRACT
is is an extended abstract for an invited keynote presentation at
the 7th Workshop on Evolutionary Computation for the Automated
Design of Algorithms (ECADA). We rst outline the motivation,
primary mechanisms, and prior results of the evolutionary com-
putation technique called “autoconstructive evolution.” We then
briey describe a collection of recent enhancements to the tech-
nique, along with a few preliminary results of ongoing experimental
work.
CCS CONCEPTS
•Computing methodologies →Genetic programming;
Arti-
cial life;
•Soware and its engineering →Genetic program-
ming;
KEYWORDS
genetic programming, autoconstructive evolution, soware synthe-
sis
ACM Reference format:
Lee Spector and Eva Moscovici. 2017. Recent Developments in Autocon-
structive Evolution. In Proceedings of GECCO ’17 Companion, Berlin, Ger-
many, July 15-19, 2017, 3 pages.
DOI: hp://dx.doi.org/10.1145/3067695.3082058
1 AUTOCONSTRUCTIVE EVOLUTION
Autoconstructive evolution is an evolutionary computation tech-
nique that has been a topic of intermient research and develop-
ment since the year 2000 [1, 2, 11, 12, 16, 17].
Autoconstructive evolution systems can be considered to be
instances of genetic programming systems, insofar as they are em-
ployed to solve computational problems for which the solutions are
executable computer programs, and insofar as they nd solutions
by processing populations of executable computer programs that
undergo cycles of variation and selection [9].
In contrast to other genetic programming systems, however, in
an autoconstructive evolution system the evolving programs are
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for prot or commercial advantage and that copies bear this notice and the full citation
on the rst page. Copyrights for components of this work owned by others than the
author(s) must be honored. Abstracting with credit is permied. To copy otherwise, or
republish, to post on servers or to redistribute to lists, requires prior specic permission
and/or a fee. Request permissions from permissions@acm.org.
GECCO ’17 Companion, Berlin, Germany
©
2017 Copyright held by the owner/author(s). Publication rights licensed to ACM.
978-1-4503-4939-0/17/07. . . $15.00
DOI: hp://dx.doi.org/10.1145/3067695.3082058
responsible not only for solving a specied computational prob-
lem, but also for producing their own children in the evolutionary
process. is allows the methods of reproduction and variation
themselves to vary and to evolve. e way that it does so is roughly
analogous to the ways in which reproductive methods (considered
broadly to include everything from karyotype to mate selection be-
haviors) evolve in biological systems; the reproductive methods are
embodied in the organisms themselves, and may evolve along with
other features of the organisms through variation and selection.
One motivation for research on autoconstructive evolution is
the speculation that it may have the potential to solve problems
beyond the reach of ordinary genetic programming systems, be-
cause evolved methods of reproduction and variation may be more
eective than those designed by hand. However, the task faced by
an autoconstructive evolution system is clearly more dicult than
that faced by an ordinary genetic programming system, because it
must “invent” and rene its reproductive mechanisms while also
solving its target computational problem.
Prior research on autoconstructive evolution has explored a va-
riety of options for the aspects of reproduction and variation that
are under the control of the evolving programs, the ways in which
child-production is integrated into the genetic programming gener-
ational loop, and the representations and instruction sets available
to programs for the construction of children. Some of this prior
work was summarized in our paper for the 2016 Workshop on Evo-
lutionary Computation for the Automated Design of Algorithms,
which also introduced the AutoDoG system on which our current
research is based [16].
2 AUTODOG
e AutoDoG system (named for “Autoconstructive Diversication
of Genomes”) is implemented on the foundation of the PushGP
genetic programming system, which has been used for a variety of
purposes and projects unrelated to autoconstructive evolution (for
example, [3, 4, 6, 7, 15]).
AutoDoG is simply PushGP, which implements a standard gen-
erational genetic programming loop, run in a conguration that
uses only the single
autoconstruction
genetic operator in place
of the collection of hand-wrien mutation and recombination oper-
ators that are normally employed. e
autoconstruction
genetic
operator takes two parent genomes as inputs and produces a child
genome by executing the program encoded by one of the parent
genomes, with both parent genomes available as data from which
the child genome can be constructed. Details are available in [
16
],
but the following are a few of the other key features of AutoDoG
relative to previous autoconstructive evolution systems:
1154
GECCO ’17 Companion, July 15-19, 2017, Berlin, Germany Lee Spector and Eva Moscovici
•
e use of a diversity-maintaining parent selection algo-
rithm (lexicase selection [
7
,
13
]), which may help to pre-
vent the collapse of population diversity as methods for
reproduction and variation evolve.
•
e representation of both parents and children using lin-
ear (Plush) genomes for structured (Push) programs [
5
],
which permits the use of relatively simple primitive in-
structions for generating and varying children.
•
Explicit diversication constraints that prevent the survival
of children that do not produce diverse children themselves,
and are therefore unlikely to contribute to the evolution
of either problem-solving or reproductive functionality.
e 2016 workshop paper demonstrated that AutoDoG could use
autoconstruction to solve a non-trivial problem, the Replace Space
with Newline problem from our soware synthesis benchmark
suite [
6
]. It also presented evidence that the reproductive methods
used by individuals in AutoDoG populations evolve in signicant
ways as runs proceed.
3 RECENT DEVELOPMENTS
Since the 2016 workshop, AutoDoG has been enhanced in several
ways, including:
•
Several new diversication tests have been developed. Pre-
viously, a child was permied to survive if and only if
two of its own children (generated only for the purpose
of this test and then discarded) diered from their parent
by dierent, non-zero amounts. Among the new tests that
have been developed are tests that require dierences more
specically in program sizes and instruction sets, and tests
that require variation over more than one generation of
descendants.
•
e domain-specic language for genome manipulation,
which programs use to generate and vary the genomes
of their children, has been enriched. Among the new ele-
ments are high-level instructions that implement compo-
nents of the hand-wrien genetic operators that are used
in non-autoconstructive runs of PushGP, such as uniform
mutation and alternation [
14
]. Evolving programs may use
these genetic operator components, with evolved param-
eters, in evolved combinations, and in conjunction with
other genome-manipulation instructions, to construct their
children.
•
A mechanism has been added for the application of “en-
tropy,” in the form of random gene deletion, to the results
of autoconstructive production of children. Because we
want variation methods to be encoded in individuals, so
that they may evolve, we do not implement exogenous
mutations upon which the evolutionary process may come
to rely. Exogenous deletions, however, are permissible, be-
cause they cannot introduce new genetic material. Initial
experiments suggest that the presence of entropy favors
individuals that repair or at least add material to their
children, presumably because these behaviors oset the
random deletions caused by entropy. is may facilitate
the emergence of meaningful variation.
•
New mechanisms have been added to favor individuals
with relatively more diversifying methods for reproduction
and variation. AutoDoG’s diversication constraints are
binary: children that meet them are allowed to survive,
while those that do not are not. e new mechanisms
provide more discrimination among those that do survive,
using stochastic screening procedures that are applied to
the population before each parent selection event. With
one method, for example, a degree of parental similarity
is randomly chosen for each parent selection event, and
individuals that are more similar to their parents than the
chosen degree are excluded from consideration. Another
method screens on the basis not of the similarity of parent
and child, but on the similarity of the parent’s and child’s
reproductive behaviors. Yet another mechanism screens
on the basis of genealogical “age,” using a method that we
call “AMPS” for “Age-Mediated Parent Selection.” AMPS
builds on the ideas of ALPS [8] and Age-Pareto Optimiza-
tion [
10
], providing a mechanism that can be added with
lile eort to standard (parent-selection-based) genetic
programming systems. When used in conjunction with
autoconstructive evolution, AMPS can provide selection
in favor of more-diversifying reproductive methods, by
incorporating child/parent dierences into the function
that is used to compute the ages of children from those of
their parents.
4 PRELIMINARY RESULTS
Experiments with the methods described above, alone and in vari-
ous combinations, are ongoing. As of this writing, we do not have
sucient data to draw rm conclusions about the ecacy of any
of these techniques, or about the long-term prospects for autocon-
structive evolution using the ideas outlined here. However, we do
have preliminary results that exhibit some points of interest.
For example, we have found that autoconstructive evolution
can sometimes achieve high success rates even within the com-
putational limits that we normally use for non-autoconstructive
genetic programming, expressed in terms of population sizes and
generation limits. is was a surprise. We have been motivated by
the prospect that autoconstructive evolution could solve problems
not solvable by non-autoconstructive genetic programming, but
we expected it to require signicantly greater resources to solve
problems that could also be solved non-autoconstructively, since
it must invent and rene its own methods for reproduction and
variation even as it works to solve the target problem.
Prior work accorded with this expectation. We had previously
documented non-autoconstructive genetic programming solving
the Replace Space with Newlines problem in roughly 50% of runs,
with a population size of 1000 and within 300 generations [
6
]. In
[
16
] we reported that AutoDoG had also solved this problem, but
that it had done so only in 5–10% of our runs, even though we
allowed runs to continue for more generations.
Using some of the new developments described above, however,
we are now able to nd solutions to the Replace Space with Newlines
problem using autoconstructive evolution at least as frequently as
we previously could with non-autoconstructive evolution, within
1155
Recent Developments in Autoconstructive Evolution GECCO ’17 Companion, July 15-19, 2017, Berlin, Germany
the same resource constraints. For example, 75% of a recent suite
of 20 AutoDoG runs with population size 1000 succeeded within
300 generations.
We have also recently conducted 20 runs of AutoDoG on the
Mirror Image soware synthesis problem [
6
], and achieved 100%
success, with 16 runs (80%) nding solutions in fewer than 300
generations. e highest success rate within 300 generations for
non-autoconstructive evolution on this problem documented in [
6
]
was 78%.
e numbers of runs described here are too low to claim with
any certainty that autoconstructive evolution performs beer than
non-autoconstructive genetic programming, on problems that both
can solve, using the same computational resources. But these re-
sults do suggest that the performance of autoconstructive evo-
lution is at least sometimes in the same ballpark as that of non-
autoconstructive genetic programming, which is notable because
of the extra work that an autoconstructive evolution system must
do to bootstrap the evolutionary process.
Recent experiments have also provided what may be the rst
evidence that autoconstructive evolution can indeed extend the
reach of genetic programming to solve problems that it could not
otherwise solve. No previous run of non-autoconstructive genetic
programming, or of any other program synthesis method of which
we are aware, has solved the String Dierences problem described
in [
6
]. However, we have recently found a (single) solution to this
problem with AutoDoG, using some of the enhancements described
above.
ese results are preliminary and largely anecdotal, and we cau-
tion the reader against concluding anything denitive about the
power or potential of autoconstructive evolution on their basis
alone. ey do, however, serve to illustrate the kinds experiments
that we are conducting, and they provide a preview of the more
complete results that we plan to present at the 7th Workshop on Evo-
lutionary Computation for the Automated Design of Algorithms.
ACKNOWLEDGMENTS
We thank the members of the Hampshire College Computational In-
telligence Lab for helpful discussions, J. Erikson for systems support,
and Hampshire College for support for the Hampshire College In-
stitute for Computational Intelligence. is material is based upon
work supported by the National Science Foundation under Grants
No. 1617087, 1129139 and 1331283. Any opinions, ndings, and
conclusions or recommendations expressed in this publication are
those of the authors and do not necessarily reect the views of the
National Science Foundation.
REFERENCES
[1]
Kyle Harrington, Emma Tosch, Lee Spector, and Jordan Pollack. 2011. Compo-
sitional Autoconstructive Dynamics. In Unifying emes in Complex Systems
Volume VIII: Proceedings of the Eighth International Conference on Complex Systems
(New England Complex Systems Institute Series on Complexity). NECSI Knowledge
Press, 856–870.
[2]
Kyle I. Harrington, Lee Spector, Jordan B. Pollack, and Una-May O’Reilly.
2012. Autoconstructive evolution for structural problems. In Proceedings of
the fourteenth international conference on Genetic and evolutionary computa-
tion conference companion (GECCO Companion ’12). ACM, 75–82.
DOI:
hp:
//dx.doi.org/10.1145/2330784.2330797
[3]
omas Helmuth. 2015. General Program Synthesis from Examples Using Genetic
Programming with Parent Selection Based on Random Lexicographic Orderings
of Test Cases. Ph.D. Dissertation. College of Information and Computer Sci-
ences, University of Massachuses Amherst, USA. hps://web.cs.umass.edu/
publication/details.php?id=2398
[4]
omas Helmuth, Nicholas Freitag McPhee, and Lee Spector. 2016. e Impact
of Hyperselection on Lexicase Selection. In GECCO ’16: Proceedings of the 2016
Annual Conference on Genetic and Evolutionary Computation. ACM, Denver, USA,
717–724.
DOI:
hp://dx.doi.org/doi:10.1145/2908812.2908851 Nominated for best
paper.
[5]
omas Helmuth, Nicholas Freitag McPhee, and Lee Spector. 2016. Plush: Linear
Genomes for Structured Push Programs. In Genetic Programming eory and
Practice XIV. Springer.
[6]
omas Helmuth and Lee Spector. 2015. General program synthesis bench-
mark suite. In GECCO ’15: Proceedings of the 2015 Conference on Genetic and
Evolutionary Computation (July, 2015).
[7]
omas Helmuth, Lee Spector, and James Matheson. 2015. Solving Uncom-
promising Problems With Lexicase Selection. Evolutionary Computation, IEEE
Transactions on 19, 5 (Oct 2015), 630–643.
DOI:
hp://dx.doi.org/10.1109/TEVC.
2014.2362729
[8]
Gregory S. Hornby. 2006. ALPS: the age-layered population structure for reducing
the problem of premature convergence. In GECCO 2006: Proceedings of the 8th
annual conference on Genetic and evolutionary computation, Vol. 1. ACM Press,
Seale, Washington, USA, 815–822. DOI:hp://dx.doi.org/doi:10.1145/1143997.
1144142
[9]
John R. Koza. 1992. Genetic Programming: On the Programming of Computers by
Means of Natural Selection. MIT Press, Cambridge, MA, USA.
[10]
Michael Schmidt and Hod Lipson. 2010. Age-Fitness Pareto Optimization. In
Genetic Programming eory and Practice VIII, Rick Riolo, Trent McConaghy,
and Ekaterina Vladislavleva (Eds.). Genetic and Evolutionary Computation,
Vol. 8. Springer, Ann Arbor, USA, Chapter 8, 129–146. hp://www.springer.
com/computer/ai/book/978-1- 4419-7746-5
[11]
Lee Spector. 2001. Autoconstructive Evolution: Push, PushGP, and Pushpop. In
Proceedings of the Genetic and Evolutionary Computation Conference (GECCO-
2001). Morgan Kaufmann, 137–146. hp://hampshire.edu/lspector/pubs/ace.pdf
[12]
Lee Spector. 2011. Towards Practical Autoconstructive Evolution: Self-Evolution
of Problem-Solving Genetic Programming Systems. In Genetic Programming e-
ory and Practice VIII, Rick Riolo, Trent McConaghy, and Ekaterina Vladislavleva
(Eds.). Genetic and Evolutionary Computation, Vol. 8. Springer New York, 17–33.
DOI:hp://dx.doi.org/10.1007/978-1- 4419-7747- 2 2
[13]
Lee Spector. 2012. Assessment of Problem Modality by Dierential Performance
of Lexicase Selection in Genetic Programming: A Preliminary Report. In 1st
workshop on Understanding Problems (GECCO-UP). ACM, 401–408.
DOI:
hp:
//dx.doi.org/doi:10.1145/2330784.2330846
[14]
Lee Spector and omas Helmuth. 2013. Uniform Linear Transformation with
Repair and Alternation in Genetic Programming. In Genetic Programming eory
and Practice XI, Rick Riolo, Jason H. Moore, and Mark Kotanchek (Eds.). Springer,
Chapter 8, 137–153. DOI:hp://dx.doi.org/doi:10.1007/978-1- 4939-0375- 7 8
[15]
Lee Spector, Jon Klein, and Maarten Keijzer. 2005. e Push3 execution stack and
the evolution of control. In GECCO 2005: Proceedings of the 2005 conference on
Genetic and evolutionary computation, Vol. 2. ACM Press, Washington DC, USA,
1689–1696. hp://www.cs.bham.ac.uk/∼wbl/biblio/gecco2005/docs/p1689.pdf
[16]
Lee Spector, Nicholas Freitag McPhee, omas Helmuth, Maggie M. Casale,
and Julian Oks. 2016. Evolution Evolves with Autoconstruction. In GECCO
’16 Companion: Proceedings of the Companion Publication of the 2016 Annual
Conference on Genetic and Evolutionary Computation. ACM, Denver, Colorado,
USA, 1349–1356. DOI:hp://dx.doi.org/doi:10.1145/2908961.2931727
[17]
Lee Spector and Alan Robinson. 2002. Genetic Programming and Autoconstruc-
tive Evolution with the Push Programming Language. Genetic Programming and
Evolvable Machines 3, 1 (March 2002), 7–40.
DOI:
hp://dx.doi.org/doi:10.1023/A:
1014538503543
1156
