You are on page 1of 6

Emphasizing Extinction in Evolutionary Programming

Garrison W. Greenwood Gary B. Fogel Manuel Ciobanu


Dept. of Computer Science Natural Selection, Inc. Dept. of Computer Science
Western Michigan University 3333 N. Torrey Pines Ct., Suite 200 Western Michigan University
Kalamazoo, MI 49008 La Jolla, CA 92037 Kalamazoo, MI 49008
garry.greenwood@wmich.edu gfogel@natural-selection.com ciobanum@cs.wmich.edu

Abstract- the highest scores are selected as parents for the next
Evolutionary programming typically uses tournament generation and the remainder are discarded [1].
selection to choose parents for reproduction. Tourna- There are several alternative views to the selection
ments naturally emphasize survival. However, a natural process in general, and tournament selection in particu-
opposite of survival is extinction, and a study of the fos- lar. Suppose we assume each individual in the popula-
sil record shows extinction plays a key role in the evolu- tion represents a distinct species. The selection process
tionary process. This paper presents a new evolutionary chooses some individuals in the current population to
algorithm that emphasizes extinction to conduct search be parents while others will be discarded. Under the
operations over a tness landscape. species abstraction, we can say that selected individuals
represent species that survive, while discarded individ-
1 Introduction uals represent species that become extinct. It has been
shown that extinction plays a signi cant role with re-
It is now well known that evolutionary programming spect to evolution in the biological world. It is therefore
(EP) can be an e ective search technique. Although natural to ask if extinction can be exploited to increase
there are several versions of EP, they are all biologically the eciency of EP. Here we review the role of extinction
inspired and generally follow the format depicted in Fig- in the fossil record and simulate this process in the con-
ure 1. text of EP. Our preliminary results suggest that extinc-
tion can enhance the performance of EP on a multimodal
initialize
population
function.

evaluate
2 Mass Extinction and the Fossil Record
fitness
The pace of change of species in ecological systems is
closely correlated to the temporal stability of the envi-
select ronments in which they inhabit. During Earth's long his-
survivors
tory, environmental stresses have been prolonged and/or
severe enough to invoke widespread ecological instability.
randomly vary A direct result of this stress is the simultaneous extinc-
individuals
tion of many species that were not able to adapt in ways
suitable to avoid their demise. These events as recorded
in the fossil record are known as mass extinctions [2].
Figure 1: The canonical EP. The \select survivors" block Extinction has played a major role in shaping the his-
includes the selection of parents and the generation of tory of life on Earth. When examining extinction rates
o spring. for a large portion of the fossil record, Raup and Sep-
koski initially noted two modes of extinction: a low-
This paper focuses on methods used for the selection level, continuous background extinction or a sporadic
of survivors as parents for the following generation. The mass extinction rate [3]. Although mass extinctions ac-
most common form of selection used in EP is tourna- count for only a small percentage of all extinctions, they
ment selection, in which the tness of each individual are singularly important because they remove \resilent"
in the population is compared against the tness of a incumbent groups from niches, creating \windows of op-
randomly chosen set of other individuals in the current portunity" for other species to ourish and establish an
population. A \win" is recorded for an individual each ecological niche [4].
time an individual's tness exceeds that of another in Five mass extinctions have been identi ed: terminal
the tournament set. All individuals are then ranked ac- Ordovician (440 Myr ago)1; late Devonian (365 Myr
cording to their number of wins. Those individuals with 1 Myr denotes one million years.
ago); late Permian (250 Myr ago); terminal Triassic
(215 Myr ago); and terminal Cretaceous (65 Myr 80 species kill (%)   
ago). The severity of these extinction events can be dif- 
cult to fully comprehend. For instance, it has been
70
  
........ ........ ........ ........ ........ ........ ........ ........ ........ ........ ........ ........ ......... ......... ........

60
.


.
.
.

estimated that 96% of all marine species on Earth went 


.
.
.
.


.

extinct by the end of the Permian period [5], representing 50



.
.
.
.
.

52% of all marine vertebrate and invertebrate families  


.

40
.
.
.
.


.
.

[3]. Several causes for these events have been proposed


.
.
.

30  
.
.
........ ........ ........ ........ ........ ........ ........ ........ ........ ........ ........ . .

including global cooling, changes in sea level, and the


.
. .
.
.


.

20
. .
.
.
. .
.

impact of a meteor or comet. Some have suggested that 


.


.
.


.
. .
.
.
.

10
.

there is a periodicity for mass extinctions with a mean 


.
.
.
.


. .


.


.


.


.


........ ........ ........ ........ ........ ........ ........ ....
.

period of 26 Myr [6, 7], although not all researchers agree


. .

0
.
.
. . .
.
.
.
. .
.
.
.

with this notion (see [8], pg 404-405). 104 105 106 107 108 109
Patterns and processes of species extinction and sur- mean waiting time in years
vival appear to be di erent between background and
mass extinction events. For instance, Jablonski [9] noted
that during background extinction intervals, Cretaceous Figure 2: A kill curve showing the average spacing
gastropods and bivalves with planktotrophic (plankton- between marine extinction events of speci ed intensity
feeding) larvae had longer species durations than those (adopted from [11]).
with nonplanktotrophic larvae. In contrast, larval type
was not a factor during the Cretaceous mass extinction
event. It is clear that not all species were a ected equally 3 Adding Extinction into Evolutionary Algo-
across all extinction events. (This fact is a central prin- rithms
ciple behind our incorporation of extinction into an evo- Our extinction-based evolutionary algorithm started as
lutionary algorithm.) a standard evolutionary programming algorithm: repro-
Wolfe and Upchurch [10] noted that one of the major duction performed only via mutation, a xed strategy
factors in the transition from predominantly evergreen parameter, and tournament selection to chose survivors.
to deciduous biotas around the world was correlated to It became necessary to alter this process when modeling
the cooling during the Cretaceous mass extinction event. extinction dynamics. For instance, not all members of
This event led to much higher rates of extinction for the population necessarily reproduce each generation; a
evergreens relative to deciduous plants. As Newton and tournament does not always determine survival; and co-
Laporte [2] noted, evolution a ects only a small subset of the population.
One answer to the question, `How did tem- We consider an ecosystem at time t that contains a
perate deciduous forests arise?' is, `Through species set
(t) = fs1 ; s2 ; : : : ; sk g, each of which is char-
enhanced survival in the Cretaceous extinc- acterized by a vector ~x 2 <n . These numbers repre-
tion.' Mass extinctions may have been re- sent phenotypical traits which change whenever species
sponsible for many such cases of evolutionary evolve. All species are subjected to an external environ-
replacement in the fossil record. mental stress (t), which may be severe enough to cause
some species to become extinct if their tness level is too
Using the stratigraphical ranges of 17,621 marine gen- low. The tness of each si is measured, in part, by an ob-
era over the last 600 Myr, Raup noted that the average jective function f (~x); minimum values of this function in-
species duration was 4 Myr. Using this same data, Raup crease the likelihood of survival. But tness is a relative
[11] developed a \kill curve", shown in Figure 2, that measure, which means f (~x) alone does not completely
described the average spacing between extinction events de ne tness unless the mapping function considers the
of varying intensity. For example, an extinction event in uence of other species in the same environment. We
capable of killing 30% of marine species occurs every 10 used the following mapping scheme. Let fmin t (f t ) de-
max
Myr; an extinction event capable of killing 65% of marine note the minimum (maximum) objective function value
species occurs every 100 Myr. Although many assump- of any species si 2
(t). Then the tness of si at time t
tions were used in the development of this curve, it ap- is given by
pears that extinctions may be continuously distributed
(s ; t) = + (1 , )  f (~xi ) , fmax
t
in a logarithmic fashion rather than separated into two i (1)
discrete categories. t , fmax
fmin t
where 2 [0; 1]. Such a mapping forces f (~x) 7! [ ; 1].
(The purpose of the lower bound will be discussed
shortly.) Other types of mappings could be used; the
above method was chosen because of its simplicity. low- t individuals intuitively have very small barriers|
The steps in our extinction-based evolutionary algo- even a small mutation could lead to a marked improved
rithm are shown below: phenotype. Roberts and Newman chose the species with
the lowest tness to evolve rst in their model; this is
1. Let t = 0 and randomly initialize population
(0). the same approach we have adopted.
Compute (si ; t) 8 i. There is no absolute scale of tness that can be consis-
2. Increment t and let
(t) =
(t , 1). tently applied over all environments; tnesses can there-
fore lie between zero and one without any loss in gener-
3. Randomly choose a stress (t) 2 [0; 0:96]. Make ality. Our algorithm maps all tness values to the subin-
all si 2
(t)j(si ; t) < (t) go extinct (i.e., discard terval [ ,1], where is a user selectable parameter. The
them). Let m equal the number of species that lower bound > 0 serves a useful purpose. Suppose
went extinct. = 0 so tness values can reside anywhere on the unit
interval. Furthermore, assume a very simplistic mapping
4. If (m > 0) then function is used|simple normalization of all tness val-
fMutate each species to produce multiple new ues to the maximum tness in the current population.
species. Compute (; t) for all new species. Con- If the evolutionary algorithm is working properly, the
duct a tournament among the new species and de- individuals should eventually evolve to a high state of
terministically choose the m best as replacements tness, which means all tness values  1. This makes
for those species that went extinct.g it very unlikely that an external stress can force species
else to become extinct|a behavior that directly contradicts
f Let j = argmin i ((si ; t)). Mutate sj . Randomly evidence found in the fossil record. The parameter cre-
choose 5 more individuals from the current popu- ates a lower threshold for extinction that prevents mass
lation and mutate them also. Compute (; t) for extinctions from occurring too frequently, while at the
all mutated species. g same time not dispensing with them entirely.
5. If computation time not exhausted, go to step 2. Reproduction is done only via mutation:
Otherwise, exit.
x0i = xi + N(0,) ; xi 2 ~x
During each generation t, three steps take place. A
random stress (t) is applied against all species, and all where N(0,) is a normally distributed random variable
species si with tness (si ; t) < (t) are discarded to with zero mean and standard deviation .
simulate extinction. At this point, one generation of a
standard EP is conducted to nd replacement species.
By restricting the replacements to be selected only from
the new species, we model speciation. However, it is
entirely possible that all of the species may have su-
ciently high tness so that none go extinct. In this case 3
the poorest t si |i.e., the species with the lowest objec-
tive function value|is mutated, which simulates a low- f 2
1
level, background evolution in
(t). Whenever a species 1
evolves to some new state, it also a ects the state and 0.5
0.
tness other surrounding species. We simulate this co- 0 0 x1
evolution by also mutating ve other randomly chosen 1
species. 0.5
0
-0.5
The stress value (t) is chosen from a uniform distri- -0.5
bution over the interval [0,0.96]. All species see the exact x2 -11
-1
same stress level. The upper limit of 0.96 sets the maxi-
mum percent of species that can go extinct in any single Figure 3: The tness landscape de ned by Eq. (2). The
generation; it is close to extinction levels seen during the global minimum is at (0,0).
Permian mass extinction [5].
There is a rationale for choosing that species with the We used our extinction-based evolutionary algorithm
lowest tness as the rst to evolve. Roberts and Newman to search for the global minimum of the multimodal ob-
[12] attached a \barrier" to all species that must be over- jective function
come in order to evolve to a state of higher tness. With
respect to an evolutionary algorithm, this barrier is a re- f (x1 ; x2 ) = x21 + 2x22 , 0:3 cos(3x1 )
gion of relatively low tness and shows the mutation size ,0:4 cos(4x2 ) + 0:7 (2)
necessary to improve phenotypical behavior. Extremely
with x1 ; x2 2 [,1; 1]. This function is shown in Figure relationship. But there is no obvious|let alone direct|
3. The algorithm was run for 10,000 generations
p with relationship between antelope and hippos nor between
N = 100 species and = 0:8. We chose  = 0:02 and lions and hippos. Does it therefore make any sense to
no self-adaptation was used. compare their phenotypes (equivalently, tnesses)? Yet,
The performance of our algorithm is compared that is exactly what happens during conventional tour-
against a standard EP that was run with the same pop- nament selection.
ulation size; the same initial population; a tournament Our extinction-based evolutionary algorithm at-
size of 20; and an identical method of reproduction. tempts to evolve a population to higher tness values
Our algorithm executes 10,000 generations about three by emulating dynamics derived from the fossil record|
times faster than a standard EP executes 5,000 gener- extinction must be present and tournaments are con-
ations. Nevertheless, our extinction-based evolutionary ducted among species. Although the exact cause of bi-
algorithm consistently outperformed the standard EP in otic extinction remains debatable, a number of reasons
searches conducted over the landscape depicted in Figure have been proposed including excessive inbreeding [13]
3. The results are given in Table 1. or environmental stresses such as climatic change, vol-
canic activity, or disease [14]. Our algorithm assumes
EA (w/extinction) standard EP environmental stress is culpable. Replacement species
minimum f (~x) 9.14010,7 6.70010,5 are created by mutating existing species|emulating
mean 8.29410 , 6 4.49810,4 speciation|and a tournament among those new candi-
std. dev. 7.95310 , 6 3.54610,4 dates determines the new species that survives. This
follows Gause's Principle of Competitive Exclusion [15]
Table 1: Performance comparison between the evolution- where new species must compete for the same ecological
ary algorithm with extinction and a standard EP. Both niche. Note that this tournament does include existing
algorithms used the same initial population to conduct species in the tournament set even though replacements
searches in the landscape shown in Figure 3. The glob- can only come from speciation.
ally minimum value is f (0; 0) = 0. Table entries are
based on data recorded from 20 independent runs.
Extinctions (%)
100
4 Discussion 80
.

The material from the two previous sections raises some


.

60
.
.
.
.
.
.
.
.

important philosophical issues regarding tournament se-


. .
. .
.
. .
. .
.
.
. .
. .
.
.
. .
. .
.

40
.
. .
. .
.

lection. In order to compute the number of wins for a


.
. .
. .
. .
.
.
. .
. .
. .
.
.
.
. .
.
. .
. .
.
. . .
.
. .
.
.
. .
. .
.
. .
. . . .
.
. . . . . . .
. . .

target individual, a small tournament set is randomly


.
. .
. .
. . .
. . .
. . . . .
. .
. ...
.
. . . . . . . . . .
. .
. .. ..
. . . . . . . .
. .. . . . . . . . .
. . . ..

20
. . . . . . . . . . . . . . .
.
.
.
. .
. .
.
. .
.
..
. .
. .
..
. .
. .
. .
. . . .
. ..
. . . .
. ..
. . .
. . .
.. ..
. . .
. .
.
.
. .
. ... ... ...... ...
. . . . . . . . . . . . . . .... ... .
. . .
. . . . . .
. .
. . . . . . .
. .
. .
. ..
.
. ........ ...

chosen from the current population. Fitness values are


. . . . ..
. . .
.. . . . . . . . .. . . . ...
... . .. .
.
.
. .
. . . .
. .
.
. .
.
..
.
. .
.
. .
.
. .
. .
. .
.
. .
. . .
. . .. . . . . . ..
. .
. .
. .
. . . ... ... ...... ....
.
.
.
. .
. . .
.
. .. . .
.. .. .. . . ..
..
. .. . .. .
.. .
. . .
. .. . .. . .
....
. . .
... . .
. .
.. ..
.
. ..
. .
.
..
.
.
.... .. ...
. . .
.... . .
. .. . . . ..
. .
. ..
. . . . . . . . .. .
.. . .
. ...... ....
.
.
.
.
.
. .
. .
.
.
.
. .
. ..
.
. ...
..
. . .
.. . .
. .... .
.
.
.
. .
.
.
.
..
. . ..
.
.
..
.
. .
.
.
.
.
.
. .
.
.
.
. ..
..
...
.
. .
.
..
. ...
.
. .
..
.
.
. .
. .
. .
..
.
..
.
.
.
.
.
..
.
.
.
...
. .... ........
.
........ ...
. . .
.... . . .. . . . . ..
. . ..
. .. . . . . . . . . . .. .
.... . . . .

then compared in a head-to-head manner. This pre-


. . . . . .
....
..
.. .. . .
...
... . . ... . .
.. . . .
.. .. ..
. ..
.. .. ....... .
.. . .
... ..
. .
.. . ........ ...........

0
.
. . . . . . .... .. . . . . . .. . . .. . .. ..
. . . . . . .
. . ..
. .
. .
.
. .
. . . ..
.
... ..
..
. .
.
.
.
..
..
..
..
..
..
...
. ......
.. .
..
.
... . . ... .
. .
...
.. . ..
. . . .
..
.
.
.
.
.
...
.
....
. ..
. .. .
..
. ..
..
....
..
...
...
.. . . . .
..... .
. .
. ..
..
. .
...
... .
... .
.. . .
..
.
..
. .
.. .
.
...
.
...
.. . . . . .
.
.......... ...... ................ ...... ......
. . .
. .
. .
.
. ..
.. .
..
..
. .
.
. .
.
. .
..
. .
..
. .
. . .
..
..
....
..
. .
.
. .
.
... ..
.
. .
.
.. .
. .
.
. ..
. .
. .
.
. .
....
..
. .
.
..
.
. .
.. ..
. .. .
..
...
..
. .
. .
...
. .. .
.. .
...
..
..
. .
.. ..
. .
..
. .
. .
.
. .
.
..
. .
. .... .... ..... ...... ......

sumes it is both acceptable and appropriate to compare 7000 7200 7400 7600 7800 8000
the tness values of distinct individuals. Tournament Generations
selection makes sense in light of a simulation at the indi-
vidual level, but di erent approaches are required when
modeling evolution at the species level. More speci cally, Figure 4: Extinctions per generation over a portion of
it may not make sense to conduct head-to-head tness the evolutionary process.
comparisons when individuals in a population are inter-
preted to be distinct species. Indeed, inter-species tness Over 25 years ago, Eldredge and Gould [16] sparked
comparison are valid only in special circumstances. a controversy that continues even today. They viewed
For clarity, let
represent the set of all species that \phyletic gradualism" as a awed concept that should be
co-exist in a given environment. The phenotype of each replaced by punctuated equilibrium. This concept says
si 2
ultimately re ects the ability of that species to that evolutionary activity is characterized by periods of
survive in that environment|it is a measure of its t- stasis, with little speciation, interspersed with episodes
ness. But, comparing the phenotypes of two distinct of very intense speciation. Figure 4 shows that our al-
species may not only be dicult, it may be meaningless. gorithm has periods of stasis interspersed with burst of
For example, in the Caprivi Game Park in northeast- extinctions. However, this behavior is not exactly punc-
ern Namibia, one can nd lions, antelopes, and hippos. tuated equilibrium|we immediately replace any extinct
These species all co-exist but do not necessarily have species whereas punctuated equilibrium does not man-
any direct relationship between each other. Of course date that a xed number of species be constantly main-
lions and antelope will since they have a predator-prey tained.
5 Conclusions & Future Work a ect distinct species in di erent ways. One obvi-
ous way is to let stress be analogous to a tness
Evolutionary change cannot be completely explained by penalty and let culling be done by natural selec-
natural selection alone [17]. At the species level, mass tion.
extinctions wipe out entire species for reasons unrelated
to inter-species struggles for survival. Nevertheless, this 2. Self-adaptive extinction
does create windows of opportunity for other species to The strategy parameter sets a lower bound on
ourish. Our algorithm follows a similar vein to conduct the percent of a species that can go extinct. As
searches: episodic mass extinctions prevent stagnation, ! 0, a large number of mass extinctions be-
and rapid speciation maintains a constant number of so- come possible|our algorithm then approximates
lutions. a (; ) evolution strategy. Conversely, as ! 1,
Our results indicate that extinction can enhance the virtually no extinctions occur and our algorithm
performance of an evolutionary search. Granted, only degenerates into a random walk with replacement.
one multimodal landscape was explored, but the re- It would be interesting to investigate the perfor-
sults are certainly encouraging particularly in light of mance when is allowed to evolve.
the low execution time with respect to a standard EP.
The improved execution time is not unexpected. In 3. Self-adaptive mutation
the extinction-based algorithm a tournament is not con- The current version of our algorithm used a xed
ducted each generation|only background evolution is standard deviation () for the normal distribution.
simulated. In fact, a tournament is conducted only after Self-adaptation should improve the performance,
the onset of a mass extinction. On the other hand, the but no single method is universally applicable [19].
standard EP conducts a tournament every generation.
Our algorithm can easily accommodate searches over 4. Other mutation distributions
dynamic landscapes. This will be the subject of future A number of recent studies indicate that alter-
studies. We have identi ed some other areas that can native probability distributions for mutation can
deserve further investigation. Speci cally, improve the performance over some landscapes
1. Modeling evolution [20, 21]. It would be a trivial matter to incorporate
It is important to note that our algorithm, in its Cauchy mutations into our algorithm.
present form, is not intended to serve as a model
for biotic macro-evolution. Nevertheless, it does Acknowledgement
incorporate a number of intriguing features that The authors would like to thank L.J. Fogel and D.B.
suggest it could be easily modi ed to perform that Fogel for their valuable comments that helped to improve
function. A rst step would be to monitor the this paper.
branching and termination of species. Ideally, we
should not maintain a constant population size|
speciation should be allowed to proceed in a nat- Bibliography
ural way with, perhaps, a maximum number of [1] T. Blickle. Tournament selection, in Handbook of
species that can be supported within the environ- Evolutionary Computation, T. Back, D. B. Fogel,
ment. and Z. Michalewicz (eds.). pages C2.3:1{4, 1997.
Normally a species can expect to undergo a num-
ber of stresses during its lifetime and all species [2] C. Newton and L. Laporte. Ancient Environments.
co-existing in that same environment will be sub- Prentice Hall, New Jersey, 3rd edition, 1989.
jected to the same stress|but not all species are [3] D. Raup and J. Sepkoski. Mass extinctions in the
a ected to the same degree by that stress. New- marine fossil record. Science, 215:1501{1503, 1982.
man [18] put it this way:
: : : we have assumed that every stress on [4] C. Marshall. Mass extinction probed. Nature,
the system a ects every species. This 392:17{20, 1998.
is clearly not realistic. Some stresses [5] D. Raup. Biological extinction in earth history. Sci-
will for example be localized in space, ence, 231:1528{1533, 1986.
or will not reach under the sea, or will
only reach under the sea, and so for. [6] D. Raup and J. Sepkowski. Periodicity of extinc-
tions in the geological past. Proc. Nat. Acad. Sci.
The current version applies the stress (t) equally USA, 81:801{805, 1984.
to all species. We could specify|or even evolve|
species relationships that permits a stress factor to
[7] ibid. Testing for periodicity of extinction. Science,
241:94{96, 1988.
[8] V. Grant. The Evolutionary Process. Columbia
University Press, NY, 2nd edition, 1991.
[9] D. Jablonski. Background and mass extinctions: the
alternation of macro-evolutionary regimes. Science,
231:129{133, 1986.
[10] J. Wolfe and G. Upchurch. Vegetation, climatic and
oral changes at the Cretaceous-Tertiary boundary.
Nature, 324:124{152, 1986.
[11] D. Raup. Extinction: Bad Genes or Bad Luck? W.
W. Norton, 1991.
[12] B. Roberts and M. Newman. A model for evolution
and extinction. J. Theor. Biol., 180:39{54, 1996.
[13] R. Frankham and K. Ralls. Inbreeding leads to ex-
tinction. Nature, 392:441{442, 1998.
[14] P. Skelton (editor). Evolution: A Biological and
Paleontological Approach. Addison-Wesley, Wok-
ingham, England, 1993.
[15] G. F. Gause. The Struggle for Existence. Williams
and Wilkins, Baltimore, 1934.
[16] N. Eldredge and S. Gould. Punctuated equilibrium:
an alternative to phyletic gradualism, in Models
in Paleobiology, T. J. Schopf (ed.), Freeman, San
Francisco. 1972.
[17] S. Gould. The evolution of life on the earth. Sci.
Amer., 271:84{91, 1994.
[18] M. E. Newman. Self-organized criticality, evolution
and the fossil extinction record. Proc. R. Soc. Lond.
B, 263:1605{1610, 1996.
[19] N. Saravanan, D. B. Fogel, and K. Nelson. A com-
parison of methods for self-adaptation in evolution-
ary algorithms. BioSys., 36:157{166, 1995.
[20] K. Chellapilla and D. B. Fogel. Two new muta-
tion operators for enhanced search and optimiza-
tion in evolutionary programming, in Applications
of Soft Computing, B. Bosacchi, J. C. Bezdek, and
D. B. Fogel (eds.). Proc. SPIE, 3165:260{269, 1997.
[21] X. Yao and Y. Liu. Fast evolution strategies. Cont.
& Cyber., 26:467{496, 1997.

You might also like