Genetic Algorithms

Author(s): John H. Holland

Source: Scientific American , Vol. 267, No. 1 (JULY 1992), pp. 66-73
Published by: Scientific American, a division of Nature America, Inc.
Stable URL: https://www.jstor.org/stable/10.2307/24939139

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide
range of content in a trusted digital archive. We use information technology and tools to increase productivity and
facilitate new forms of scholarship. For more information about JSTOR, please contact [email protected].

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
https://about.jstor.org/terms

Scientific American, a division of Nature America, Inc. is collaborating with JSTOR to digitize,
preserve and extend access to Scientific American

This content downloaded from

192.188.53.214 on Tue, 04 Jul 2023 15:50:28 +00:00
All use subject to https://about.jstor.org/terms
 Genetic Algorithms
Computer programs that ({evolve" in ways that resemble
natural selection can solve complex problems
even their creators do not fully understand
by John H. Holland

n ·ving organisms are consummate

problem solvers. They exhibit a ver­
derstood conditions, the practical re­
sults they achieve may yield some in­
and early 1960s, fared poorly because
they followed the emphasis in biolOgi­
satility that puts the best com­ sight into the details of how life and in­ cal texts of the time and relied on muta­
puter programs to shame. This observa­ telligence evolved in the natural world. tion rather than mating to generate new
tion is especially galling for computer Most organisms evolve by means of gene combinations. Then, in the early
scientists, who may spend months or two primary processes: natural selec­ 1960s, Hans]. Bremermann of the Uni­
years of intellectual effort on an al­ tion and sexual reproduction. The first versity of California at Berkeley added
gorithm, whereas organisms come by determines which members of a pop­ a kind of mating: the characteristics of
their abilities through the apparently ulation survive to reproduce, and the offspring were determined by summing
undirected mechanism of evolution and second ensures mixing and recombina­ up corresponding genes in the two par­
natural selection. tion among the genes of their offspring. ents. This mating procedure was limit­
Pragmatic researchers see evolution's When sperm and ova fuse, matching ed, however, because it could apply only
remarkable power as something to be chromosomes line up with one another to characteristics that could be added
emulated rather than envied. Natural and then cross over partway along their together in a meaningful way.
selection eliminates one of the greatest length, thus swapping genetic material.

D
hurdles in software design: specifying This mixing allows creatures to evolve uring this time, I had been in­
in advance all the features of a problem much more rapidly than they would if vestigating mathematical analy­
and the actions a program should take each offspring simply contained a copy ses of adaptation and had be­
to deal with them. By harnessing the of the genes of a single parent, modified come convinced that the recombination
mechanisms of evolution, researchers occasionally by mutation. (Although uni­ of groups of genes by means of mating
may be able to "breed" programs that cellular organisms do not engage in was a critical part of evolution. By the
solve problems even when no person mating as humans like to think of it, mid-1960s I had developed a program­
can fully understand their structure. In­ they do exchange genetic material, and ming technique, the genetic algorithm,
deed, these so-called genetic algorithms their evolution can be described in anal­ that is well suited to evolution by both
have already demonstrated the ability ogous terms.) mating and mutation. During the next
to make breakthroughs in the design of Selection is simple: if an organism decade, I worked to extend the scope of
such complex systems as jet engines. fails some test of fitness, such as recog­ genetic algorithms by creating a genet­
Genetic algorithms make it possible nizing a predator and fleeing, it dies. ic code that could represent the struc­
to explore a far greater range of poten­ Similarly, computer scientists have little ture of any computer program.
tial solutions to a problem than do con­ trouble weeding out poorly performing The result was the classifier sys­
ventional programs. Furthermore, as re­ algorithms. If a program is supposed to tem, consisting of a set of rules, each
searchers probe the natural selection of sort numbers in ascending order, for ex­ of which performs particular actions
programs under controlled and well-un- ample, one need merely check whether every time its conditions are satisfied
each entry of the program's output is by some piece of information. The con­
larger than the preceding one. ditions and actions are represented by
JOHN H. HOLLAND has been investi­ People have employed a combination strings of bits corresponding to the
gating the theory and practice of algo­ of crossbreeding and selection for mil­ presence or absence of specific charac­
rithmic evolution for nearly 40 years. He lennia to breed better crops, racehorses teristics in the rules' input and output.
is a professor of psychology and of elec­
or ornamental roses. It is not as easy, For each characteristic that was pres­
trical engineering and computer science
however, to translate these procedures ent, the string would contain a 1 in the
at the University of Michigan. Holland
received a B.S. in physics from the Mas­ for use on computer programs. The appropriate position, and for each that
sachusetts Institute of Technology in chief problem is the construction of a was absent, it would contain a O. For ex­
1950 and served on the Logical Planning "genetic code" that can represent the ample, a classifier rule that recognized
Group for IBM's first programmed elec­ structure of different programs, just as dogs might be encoded as a string con­
tronic computer (the 701) from 1950 un­ DNA represents the structure of a per­ taining l's for the bits corresponding to
til 1952. He received an M.A. in math·
son or a mouse. Mating or mutating the "hairy," "slobbers," "barks," "loyal" and
ematics and a Ph.D. in communication
text of a FORTRAN program, for exam­ "chases sticks" and O's for the bits cor­
SCiences from the University of Michi­
gan. Holland has been a member of the
ple, would in most cases not produce a responding to "metallic," "speaks Urdu"
Steering Committee of the Santa Fe in­ better or worse FORTRAN program but and "possesses credit cards." More real­
stitute since its inception in 1987 and is rather no program at all. istically, the programmer should choose
an external professor there. The first attempts to mesh computer the Simplest, most primitive character­
science and evolution, in the late 1950s istics so that they can be combined-as

66 SCIENTIFIC AMERICAN July 1992

© 1992 SCIENTIFIC AMERICAN, INC

This content downloaded from

192.188.53.214 on Tue, 04 Jul 2023 15:50:28 +00:00
All use subject to https://about.jstor.org/terms
in the game of 20 Questions-to classify Furthermore, the mating process con­ kind of binary equivalent of map coor­
a wide range of objects and situations. tinually combines these strings in new dinates. The set of all strings that start
Although they excel at recognition, ways, generating ever more sophisti­ with a 1, for example, constitutes a re­
these rules can also be made to trigger cated solutions. The kinds of problems gion in the set of possibilities. So do all
actions by tying bits in their output to that have yielded to the technique range the strings that start with a 0 or that
the appropriate behavior [see illustra­ from developing novel strategies in have a 1 in the fourth position, a 0 in
tion on page 69]. Any program that can game theory to designing complex me­ the fifth and a 1 in the sixth and so on.
be written in a standard programming chanical systems. One conventional technique for ex­
language such as FORTRAN or llSP can ploring such a landscape is hill climb­

R
be rewritten as a classifier system. ecast in the language of genetic ing: start at some random point, and if
To evolve classifier rules that solve a algorithms, the search for a good a slight modification improves the qual­
particular problem, one simply starts solution to a problem is a search ity of your solution, continue in that di­
with a population of random strings of for particular binary strings. The uni­ rection; otherwise, go in the opposite
l's and O's and rates each string ac­ verse of all possible strings can be con­ direction. Complex problems, howev­
cording to the quality of its result. De­ sidered as an imaginary landscape; val­ er, make landscapes with many high
pending on the problem, the measure leys mark the location of strings that points. As the number of dimensions of
of fitness could be business profitabili­ encode poor solutions, and the land­ the problem space increases, the coun­
ty, game payoff, error rate or any num­ scape's highest point corresponds to tryside may contain tunnels, bridges
ber of other criteria. High-quality strings the best possible string. and even more convoluted topological
mate; low-quality ones perish. As gen­ Regions in the solution space can also features. Finding the right hill or even
erations pass, strings associated with be defined by looking at strings that determining which way is up becomes
improved solutions will predominate. have l's or O's in speCified places-a increasingly difficult. In addition, such

BEE ORCHID demonstrates the specificity with which natural to natural selection, the author says, can produce computer
genetic selection can match an organism to a particular niche. programs (so-called genetic algorithms) capable of solving
The flower, which resembles a female bumblebee, is fertilized such complex problems as the design of jet turbines or com­
by male bees that attempt to mate with it. Mechanisms similar munications networks.

SCIENTIFIC AMERICAN July 1992 67

© 1992 SCIENTIFIC AMERICAN, INC

This content downloaded from

192.188.53.214 on Tue, 04 Jul 2023 15:50:28 +00:00
All use subject to https://about.jstor.org/terms
CROSSOVER is the fundamental mechanism of genetic rear­ Chromosomes line up and then swap the portions of their ge­
rangement for both real organisms and genetic algorithms. netic code beyond the crossover point.

search spaces are usually enormous. If 10,000 symbols flips from 0 to 1, or vice implicit parallelism. The purpose of
each move in a chess game, for exam­ versa. Mutation alone does not general­ crossing strings in the genetic algorithm
ple, has an average of 10 alternatives, ly advance the search for a solUtion, is to test new parts of target regions
and a typical game lasts for 30 moves but it does provide insurance against rather than testing the same string over
on each side, then there are about 1060 the development of a uniform popula­ and over again in successive genera­
strategies for playing chess (most of tion incapable of further evolution. tions. But the process can also "move"
them bad). an offspring out of one region into an­

T
Genetic algorithms cast a net over this he genetic algorithm exploits the other, causing the sampling rate of dif­
landscape. The multitude of strings in higher-payoff, or "target," regions ferent regions to depart from a strict
an evolving population samples it in of the solution space, because proportionality to average fitness. That
many regions simultaneously. Notably, successive generations of reproduction departure will slow the rate of evolution.
the rate at which the genetic algorithm and crossover produce increasing num­ The probability that the offspring of
samples different regions corresponds bers of strings in those regions. The al­ two strings will leave its parents' re­
directly to the regions' average "eleva­ gorithm favors the fittest strings as gion depends on the distance between
tionn-that is, the probability of finding parents, and so above-average strings the 1's and O's that define the region.
a good solution in that vicinity. (which fall in target regions) will have The offspring of a string that samples
This remarkable ability of genetic more offspring in the next generation. 10****, for example, can be outside
algorithms to focus their attention on Indeed, the number of strings in a giv­ that region only if crossover begins at
the most promising parts of a solution en region increases at a rate proportion­ the second position in the string-one
space is a direct outcome of their ability al to the statistical estimate of that re­ chance in five for a string containing
to combine strings containing partial so­ gion's fitness. A statistician would need six genes. (The same building block
lutions. First, each string in the popula­ to evaluate dozens of samples from would run a risk of only one in 999 if
tion is evaluated to determine the per­ thousands or millions of regions to es­ contained in a 1,000-gene string.) The
formance of the strategy that it encodes. timate the average fitness of each re­ offspring of a six-gene string that sam­
Second, the higher-ranking strings mate. gion. The genetic algorithm manages to ples region 1 ***,' 1 runs the risk of leav­
Two strings line up, a point along the achieve the same result with far fewer ing its parents' region no matter where
strings is selected at random and the strings and virtually no computation. crossover occurs.
portions to the left of that point are The key to this rather surprising be­ Closely adjacent l's or O's that define
exchanged to produce two offspring: havior is the fact that a single string a region are called compact building
one containing the symbols of the first belongs to all the regions in which any blocks. They are most likely to survive
string up to the crossover point and of its bits appear. For example, the crossover intact and so be propagated
those of the second beyond it, and the string 11011001 is a member of regions into future generations at a rate propor­
other containing the complementary 11 ****,,* (where the * indicates that a tional to the average fitness of strings
cross [see illustration above]. Biological bit's value is unspecified), 1***,,*',1, that carry them. Although a reproduc­
chromosomes cross over one another **0**00* and so forth. The largest re­ tion mechanism that includes crossover
when two gametes meet to form a zy­ gions-those containing many unspeci­ does not manage to sample all regions
gote, and so the process of crossover in fied bits-will typically be sampled by at a rate proportional to their fitness, it
genetic algorithms does in fact close­ a large fraction of all the strings in a does succeed in doing so for all regions
ly mimic its biological model. The off­ population. Thus, a genetic algorithm defined by compact building blocks.
spring do not replace the parent strings; that manipulates a population of a few The number of compactly defined build­
instead they replace low-fitness strings, thousand strings actually samples a ing blocks in a population of strings still
which are discarded at each generation vastly larger number of regions. This vastly exceeds the number of strings,
so that the total population remains the implicit parallelism gives the genetic al­ and so the genetic algorithm still ex­
same size. gorithm its central advantage over oth­ hibits implicit parallelism.
Third, mutations modify a small frac­ er problem-solving processes. Curiously, an operation in natural ge­
tion of the strings: roughly one in every Crossover complicates the effects of netics called inversion occasionally rear-

68 SCIENTIFIC AMERICAN July 1992

© 1992 SCIENTIFIC AMERICAN, INC

This content downloaded from

192.188.53.214 on Tue, 04 Jul 2023 15:50:28 +00:00
All use subject to https://about.jstor.org/terms
ranges genes so that those far apart in pIe, it is possible to concentrate on ex­ breaks up a building block, it produces
the parents may be placed close to one ploiting that strategy. But this choice a new block that enables the genetic al­
another in the offspring. This amounts carries a hidden cost because exploita­ gorithm to test regions it has not previ­
to redefining a building block so that it tion makes the discovery of truly novel ously sampled.
is more compact and less subject to be­ strategies unlikely. Improvements come The game known as the Prisoner's
ing broken up by crossover. If the build­ from trying new, risky things. Because Dilemma illustrates the genetic algo­
ing block specifies a region of high aver­ many of the risks fail, exploration in­ rithm's ability to balance exploration
age fitness, then the more compact ver­ volves a degradation of performance. against exploitation. This long-studied
sion automatically displaces the less Deciding to what degree the present game presents its two players with a
compact one because it loses fewer off­ should be mortgaged for the future is choice between "cooperation" and "de­
spring to copying error. As a result, a classic problem for all systems that fection": if both players cooperate, they
an adaptive system using inversion can adapt and learn. both receive a payoff; if one player de­
discover and favor compact versions of The genetic algorithm's approach to fects, the defector receives a higher
useful building blocks. this obstacle turns on crossover. Al­ payoff and the cooperator receives
The genetic algorithm's impliCit par­ though crossover can interfere with nothing; if both defect, they both receive
allelism allows it to test and exploit the exploitation of a building block by a minimal payoff [see table on page 71].
large numbers of regions in the search breaking it up, this process of recombi­ For example, if player A cooperates and
space while manipulating relatively few nation tests building blocks in new player B defects, then player A receives
strings. Implicit parallelism also helps combinations and new contexts. Cross­ no payoff and player B receives a pay­
genetic algorithms to cope with nonlin­ over generates new samples of above­ off of five points.
ear problems-those in which the fit­ average regions, confirming or disprov­ Political scientists and sociologists
ness of a string containing two particu­ ing the estimates produced by earlier have studied the Prisoner's Dilemma be­
lar building blocks may be much great­ samples. Furthermore, when crossover cause it provides a Simple, clear-cut ex-
er (or much smaller) than the sum of
the fitnesses attributable to each build­
ing block alone.
Linear problems present a reduced
search space because the presence of a How to Build a Classifier System
1 or a 0 at one position in a string has

B can evolve requires a way of rep­

no effect on the fitness attributable to uilding a computer algorithm that (what the program does). A classifier
the presence of a 1 or 0 somewhere
consists simply of strings represent­
resenting the program so that any ing possible characteristics of the pro­
else. The space of 1,000-digit strings,
change in its genotype (the bits that gram's input and actions to take
for example, contains more than 31,000 compose the program) leads to a (be/ow). Changing any symbol in a
possibilities, but if the problem is lin­ meaningful change in its phenotype string changes its behavior.
ear, an algorithm need investigate only
strings containing 1 or 0 at each posi­
tion, a total of just 2,000 possibilities.
When the problem is nonlinear, the
difficulty increases sharply. The average
fitness of strings in the region ,,01***, A classifier system to emulate a frog, for ex­
for example, could be above the popu­ ample, might contain strings that react to
lation average, even though the fitness­ objects that the frog sees. Depending on an
object's motion, size, location and other at­
es associated with ,,0;,*** and ** 1***
tributes, the frog would attack, pursue or ig­
are below the population average. Non­
nore it. Several strings may match the same
linearity does not mean that no useful input string; the one with the fewest "don't
building blocks exist but merely that care" symbols governs the system's actions.
blocks consisting of single l's or O's
will be inadequate. That characteristic,
in turn, leads to an explosion of possi­
bilities: the set of all strings 20 bits in
length contains more than three billion
building blocks. Fortunately, implicit
INPUT OUTPUT
parallelism can still be exploited. In a
population of a few thousand strings,
many compact building blocks will ap­
pear in 100 strings or more, enough to
provide a good statistical sample. Build­
ing blocks that exploit nonlinearities to IF OBJECT IS MOVING, FLEE
attain above-average performance will
automatically be used more often in fu­
ture generations.
In addition to coping with nonlineari­
IF OBJECT IS MOVING, IN THE AIR, SMALL AND NEAR, PURSUE
ty, the genetic algorithm helps to solve
a conundrum that has long bedeviled
conventional problem-solving methods:
striking a balance between exploration
and exploitation. Once one finds a good IF OBJECT IS MOVING, IN THE AIR, SMALL, NEAR AND STRIPED, DO NOTHING
strategy for playing chess, for exam-

SCIENTIFIC AMERICAN July 1992 69

© 1992 SCIENTIFIC AMERICAN, INC

This content downloaded from

192.188.53.214 on Tue, 04 Jul 2023 15:50:28 +00:00
All use subject to https://about.jstor.org/terms
 INITIAL POPULATION MATING INTERMEDIATE POPULATION MATING

0.60 �.J!..:;.;"-"-_____----'

0.45 L.:..:.�
.: -L..._____----'

1 110(§'bt�1 10001111100111011111 1 X

1 000101100100111100101110101111 1- __
__
__
__
__
_

__
__
_

0.30 GI �00�10�1�01�10�1�01�00�0�10�1�100�1�00�1100�1�0 1 -------------:)� 0.30 1 1001010110101000101�0010010010 1 X

0.25 1 11110011 1100100110100000000011 1 ) 0.25 1 111100111100100110100000000011 1 X

GENE POOL of algorithms consists of strings of l's and O's. Each string is evaluated
1
0.20 001111100001111011001101110010 1 X
for fitness, and the best strings mate (second column) and produce offspring by
means of crossover (indicated by a vertical black line). Strings of intermediate
0.20 1 1000100000000100101�001101qooo l X
fitness simply survive to the next generation, and the least fit perish. If particular
patterns of bits (shown here by colored areas) improve the fitness of strings that
carry them, repeated cycles of evaluation and mating (succeeding columns) will
1
0.10 110101100000101101010010100110 1 X cause the proportion of these high-quality "building blocks" to increase. The pattern
corresponding to each building block appears in the rightmost column; asterisks
0.10 1 010010� 101o o,110111110101111111 1 X represent bits whose values are unspecified.

ample of the difficulties of cooperation. whether the preferred response to its best individual.) Similarly, the genetic
Game theory predicts that each player corresponding history was cooperation algorithm can be used, with modifica­
should minimize the maximum dam­ or defection. For example, the 54-bit tions, to govern the evolution not mere­
age the other player can inflict: that is, string consisting of all O's would desig­ ly of individual rules or strategies but
both players should defect. Yet when nate the strategy that defects in all cas­ of classifier-system "organisms" com­
two people play the game together re­ es. Even for such a simple game, there posed of many rules. Instead of select­
peatedly, they typically learn to cooper­ are 264 (approximately 15 quadrillion) ing the fittest rules in isolation, com­
ate with each other to raise their joint different strategies. petitive pressures can lead to the evo­
payoff. One of the most effective known Axelrod and Forrest supplied the ge­ lution of larger systems whose abilities
strategies for the Prisoner's Dilemma is netic algorithm with a small random are encoded in the strings that make
"tit for tat," which begins by cooperat­ collection of strings representing strate­ them up.
ing but thereafter mimics the last play gies. The fitness of each string was sim­ Re-creating evolution at this higher
of the other player. That is, it "punish­ ply the average of the payoffs its strate­ level requires several modifications to
es" a defection by defecting the next gy received under repeated play. All the original genetic algorithm. Strings
time, and it rewards cooperation by co­ these strings had low fitnesses because still represent condition-action rules,
operating the next time. most strategies for playing the Prison­ and each rule whose conditions are met
Robert Axelrod of the University of er's Dilemma are not very good. Quickly generates an action as before. Rating
Michigan, working with Stephanie For­ the genetic algorithm discovered and ex­ each rule by the number of correct ac­
rest, now at the University of New Mex­ ploited tit for tat, but further evolution tions it generates, however, will favor
ico, decided to find out if the genetic introduced an additional improvement. the evolution of individual "superrules"
algorithm could discover the tit-for-tat The new strategy, discovered while the instead of finding clusters of rules that
strategy. Applying the genetic algorithm genetic algorithm was already playing at interact usefully. To redirect the search
first requires translating possible strate­ a high level, exploited players that could toward interacting rules, the procedure
gies into strings. One simple way is to be "bluffed"-lured into cooperating re­ is modified by forcing rules to compete
base the next response on the outcome peatedly in the face of defection. It re­ for control of the system's actions. Each
of the last three plays. Each iteration verted to tit for tat, however, when the rule whose conditions are met com­
has four possible outcomes, and so a history indicated the player could not petes with all other rules whose condi­
sequence of three plays yields 54 pos­ be bluffed. tions are met, and the strongest rules
sibilities. A 54-bit string contains one determine what the system will do in

B
gene (or bit position) for each. The first iological evolution operates, of that given situation. If the system's ac­
gene, for instance, would be allocated course, not to produce a single tions lead to a successful outcome, all
to the case of three consecutive mutual superindividual but rather to pro­ the winning rules are strengthened; oth­
cooperations and the last to three mu­ duce interacting species well adapted erwise they are weakened.
tual defections. The value of each gene to one another. (Indeed, in the biologi­ Another way of looking at this meth­
would be either 1 or 0 depending on cal realm there is no such thing as a od is to consider each rule string as a

70 SCIENTIFIC AMERICAN July 1992

© 1992 SCIENTIFIC AMERICAN, INC

This content downloaded from

192.188.53.214 on Tue, 04 Jul 2023 15:50:28 +00:00
All use subject to https://about.jstor.org/terms
 INTERMEDIATE POPULATION MATING FINAL POPULATION KEY

l ,0
0.60 Oll!1 1110 11001101001010001110 1
1 J
0.65 011 1111 Oll00ll0l0 l 00 l 0l l 00ll I 1........ ...... ......··· 0 1 1 1· 1
• •

1 �1
0.70 011 1110011001101010010110011 1
� 0.70 L-'--. :..:.--'..:.
.. 11000100001010001110
....""'--
_---'-'- ....:. �.:: 0.60 I 011�1111doll00ll0l00l0l000lllO l 1.......... ..1 0 0 1..1 0 ..· 1
••••••

� 0.60 I U
0.70 011 1111 O,Ol1000100001010001110 1 1......1 1.1 0 0 .. ............·· 1
•
L-________-L__��

1
�
� 0.60 101010010100011101001010001110 0.60 101010010100010100010010110011 1
1
0.55 101010010100011101001010001110 1 1 �
0.65 011 11 1 10011000100010010110011 1
� 0.55 I 1010100101000101000W010ll00ll I )( 1
0.70 101010010100010100001010001110 1
1 g1 10001111101110101111 1 )(
�0.50 110

�0.50 1 000101100100111100101110101111 1 )( 1
0.60 101010010100011101001010001110 1

hypothesis about the classifier's world. strong enough to dictate behavior in that increase pressure in a particular
A rule enters the competition only when particular cases. branch of the pipeline and valves to
it "claims" to be relevant to the current Eventually the system develops a hi­ regulate the flow of gas to and from
situation. Its ability to compete depends erarchy: layers of exception rules at the storage tanks. Because of the tremen­
on how much of a contribution it has lower levels handle most cases, but the dous lag between manipulating valves
made to solving similar problems. As default rules at the top level of the hi­ or compressors and the actual pressure
the genetic algorithm proceeds, strong erarchy come into play when none of changes in the lines, there is no analyt­
rules mate and form offspring rules the detailed rules has enough informa­ ic solution to the problem, and human
that combine their parents' building tion to satisfy its conditions. Such de­ controllers, like Goldberg's algorithm,
blocks. These offspring, which replace fault hierarchies bring relevant experi­ must learn by apprenticeship.
the weakest rules, amount to plausible ence to bear on novel situations while Goldberg's system not only met gas
but untried hypotheses. preventing the system from becoming demand at costs comparable to those
Competition among rules provides bogged down in overly detailed options. achieved in practice, but it also devel­
the system with a graceful way of han­ The same characteristics that make oped a hierarchy of default rules ca­
dling perpetual novelty. When a system evolving classifier systems adept at han­ pable of responding properly to holes
has strong rules that respond to a par­ dling perpetual novelty also do a good punched in the pipeline (as happens
ticular situation, that is the equivalent job of handling situations where the all too often in reality at the blade of
of saying that it has certain well-validat­ payoff for a given action may come an errant bulldozer). Lawrence Davis of
ed hypotheses. Offspring rules, which only long after the action is taken. The Tica Associates in Cambridge, Mass.,
begin life weaker than do their parents, earliest moves of a chess game, for ex­ has used similar techniques to design
can win the competition and influence ample, may set the stage for later victo­ communications networks; his soft­
the system's behavior only when there ry or defeat. ware's goal is to carry the maximum
are no strong rules whose conditions To train a classifier system for such possible amount of data with the mini­
are satisfied-in other words, when the long-term goals, a programmer gives mum number of transmission lines and
system does not know what to do. If the system a payoff each time it com­ switches interconnecting them.
their actions help, they survive; if not, pletes a task. The credit for success (or A group of researchers at General
they are soon replaced. Thus, the off­ the blame for failure) can propagate Electric and Rensselaer Polytechnic In-
spring do not interfere with the sys­ through the hierarchy to strengthen (or
tem's action in well-practiced situations weaken) individual rules even if their
but wait gracefully in the wings as hy­ actions had only a distant effect on the
potheses about what to do under novel outcome. Over the course of many gen­ The Prisoner's Dilemma
circumstances. erations the system develops rules that
(B) (B)
Adding competition in this way act ever earlier to set the stage for later PLAYER
COOPERATE DEFECT
strongly affects the evolution of a clas­ payoffs. It therefore becomes increas­
sifier system. Shortly after the system ingly able to anticipate the consequenc­
(A)
starts running, it evolves rules with sim­ es of its actions. 3/3 5/0
COOPERATE
ple conditions-treating a broad range

G
of situations as if they were identical. enetic algorithms have now been
(A)
The system exploits such rules as de­ tested in a wide variety of con­ 0/5 0/0
DEFECT
faults that specify something to be done texts. David E. Goldberg of the
in the absence of more detailed infor­ University of Illinois, for example, has
mation. Because the default rules make developed algorithms that learn to con­
IN PRISONER'S DILEMMA each play­
only coarse discriminations, however, trol a gas pipeline system modeled on er can either cooperate or defect and
they are often wrong and so do not the one that carries natural gas from the receives a payoff based on the other's
grow in strength. As the system gains Southwest to the Northeast. The pipe­ choice. If both cooperate, for example,
experience, reproduction and crossover line complex consists of many branches, both receive three points. Mutual defec­
lead to the development of more com­ all carrying various amounts of gas; the tion is the safest strategy, but repeated
plex, speCific rules that rapidly become only controls available are compressors play often leads to cooperation instead.

SCIENTIFIC AMERICAN July 1992 71

© 1992 SCIENTIFIC AMERICAN, INC

This content downloaded from

192.188.53.214 on Tue, 04 Jul 2023 15:50:28 +00:00
All use subject to https://about.jstor.org/terms
 the advantages of implicit parallelism.
As maSSively parallel computers be­
come more common, it will be feasible
to evolve software populations whose
size more closely approaches those of
natural species. Indeed, the genetic al­
gorithm lends itself nicely to such ma­
chines. Each processor can be devoted to
a single string because the algorithm's
operations focus on single strings or,
at most, a pair of strings during cross­
over. As a result, the entire population
can be processed in parallel.
We still have much to learn about
classifier systems, but the work done
so far suggests they will be capable of
remarkably complex behavior. Rick L.
Riolo of the University of Michigan has
already observed genetic algorithms
that display "latent learning" (a phe­
nomenon in which an animal such as a
rat explores a maze without reward and
is subsequently able to find food placed
SOFlWARE TO DESIGN JET TURBINE includes a genetic algorithm that combines in the maze much more quickly).
the best features of designs produced by other programs. Engineers using the algo­ At the Santa Fe Institute, Forrest, W.
rithm achieved better results than with more conventional software aids. Brian Arthur, John H. Miller, Richard G.
Palmer and I have used classifier sys­
tems to simulate economic agents of
stitute recently put a genetic algorithm ly. These dead ends occur because it is limited rationality. These agents evolve
to good use in the design of a high-by­ practically impossible to sort out all the to the point of acting on trends in a
pass jet engine turbine such as those effects associated with different multi­ simple commodity market, producing
that power commercial airliners. Such ple changes, let alone to specify the re­ speculative bubbles and crashes.
turbines, which consist of multiple stag­ gions of the design space within which The simulated worlds these agents
es of stationary and rotating blade rows previous experience remains valid. inhabit show many similarities to natu­
enclosed in a roughly cylindrical duct, To get away from such a point, the ral ecosystems: they exhibit counter­
are at the center of engine-development designer must find new building blocks parts to such phenomena as symbio­
projects that last five years or more and for a solution. Here is where the genet­ sis, parasitism, biological "arms races,"
consume up to $ 2 billion. ic algorithm comes into play. Seeding mimicry, niche formation and specia­
The design of a turbine involves at the algorithm with designs produced tion. Still other work with genetic algo­
least 100 variables, each of which can by the expert system, an engineer took rithms has shed light on the conditions
take on a different range of values. The only two days to find a design with under which evolution will favor sexual
resulting search space contains more three times the improvements of the or asexual reproduction. Eventually ar­
than 10387 points. The "fitness" of the manual version (and half again as many tificial adaptation may repay its debt to
turbine depends on how well it satisfies as using the expert system alone). nature by increasing researchers' un­
a series of 50 or so constraints, such as This example points up both a derstanding of natural ecosystems and
the smooth shape of its inner and out­ strength and a limitation of simple ge­ other complex adaptive systems.
er walls or the pressure, velocity and netic algorithms: they are at their best
turbulence of the flow at various points when exploring complex landscapes to
inside the cylinder. Evaluating a single locate regions of enhanced opportuni­ FURTIlER READING
design requires running an engine sim­ ty. But if a partial solution can be im­ INDUCTION: PROCESSES OF INFERENCE,
ulation that takes about 30 seconds on proved further by making small chang­ LEARNING, AND DISCOVERY. J. H. Hol­
a typical engineering workstation. es in a few variables, it is best to aug­ land, K. J. Holyoak, R. E. Nisbett and P.
In one fairly typical case, an engi­ ment the genetic algorithm with other, R. Thagard. MIT Press, 1986.
GENETIC ALGORITHMS AND SIMUIATED

�
neer working alone took about eight more standard methods.
ANNEAuNG. Edited by Lawrence Davis.
weeks to reach a satisfactory design.
Morgan Kaufmann, 1987.
So-called expert systems, which use in­ thOUgh genetic algorithms mimic
GENETIC ALGORITHMS IN SEARCH, OP­
ference rules based on experience to the effects of natural selection, TIMIZATION, AND MACHINE LEARNING.
predict the effects of a change of one until now they have operated D. E. Goldberg. Addison-Wesley, 1989.
or two variables, can help direct the de­ on a much smaller scale than does bio­ GENETIC ALGORlTHMS: PROCEEDINGS OF
signer in seeking out useful changes. logical evolution. My colleagues and I THE FOURTH INTERNATIONAL CONFER­
An engineer using such an expert sys­ have run classifier systems containing ENCE. Edited by Richard Belew and
Lashon Booker. Morgan Kaufmann, 1991.
tem took less than a day to design an as many as 8,000 rules, but this size is
ADAPTATION IN NATURAL AND ARTIFI­
engine with twice the improvements of at the low end of viability for natural
CIAL SYSTEMS. J. H. Holland. MIT Press,
the eight-week manual design. populations. Large animals that are not 1992.
Such expert systems, however, soon endangered may number in the mil­ COMPLEX ADAPTIVE SYSTEMS. J. H. Hol­
get stuck at points where further lions, insect populations in the trillions land in Dazda/us, Vol. 121, No.1, pages
improvements can be made only by and bacteria in the quintillions or more. 17-30; Winter 1992.
changing many variables simultaneous- These large numbers greatly enhance

72 SCIENTIFIC AMERICAN ruly 1992

© 1992 SCIENTIFIC AMERICAN, INC

This content downloaded from

192.188.53.214 on Tue, 04 Jul 2023 15:50:28 +00:00
All use subject to https://about.jstor.org/terms
 Effortlessly contour and
colormap surfaces generated f--iT�,
from random sample pOints.

Pull a lever to adjust a variable, as the

revolutionary Graphics Equalizer� Pop up the extraordinary
animates changes to your curve-fits graph gallery of over 160
and formulas on-screen. business, technical,
and statistical graph types.

Automate your presentation

with master slides, outliner,
speaker & handout notes.

Access your numbers

instantly from your
favorite Windows
or DOS spreadsheet.

Hot-link multiple
data files directly to
Stanford Graphics'
proprietary 4-0
70-trillion
cell spreadsheet.

Instantly identify
& modify your data
points with the
Intelligent Data Cursor.�
Simply drag a bar or
line to change a cell in
your spreadsheet!

Click on the Calculator

for on-the-fly business,
financial or scientific
computations.

STANFORD GRAPHICS FOR WINDOWS

Finally... presentation graphics with numerical clout!
"Stanford Graphics offers the best of both worlds.
"Stanford is a breakthrough product that combines
It combines top-notch presentation features with
technical charting. data analysis, and presentation "Stanford Graphics slams the competition ..:'
tools. Basically, it's fabulous." *
unprecedented technical sophistication. � were
- PC Magazine greatly impressed:' _ Infoworld. - Infoworld
March. 1992 Sept.. 1991 April. 1992

Does your presentation software give in Windows 3.0 and 3.1 technology, with
you all beauty and no brains? Where is sophisticated DDE, OLE, and TrueType
the substance behind those pretty pic­ font support.
tures?What if your presentation program Stanford Graphics for Windows:
had enough data-handling muscle to Presentation and Analysis.
meet the needs of business and statistical Finally, form and substance.
and scientific users?
With on-screen curve fitting. formula 3IlL 3-� UISII±INS
solVing. statistics, and data manipulation,
Stanford Graphics advances both the art
,rCall Toll Free for Details
and science of technical presentations.
STANFORD GRAPHICS"
1-800-729-4723
All this and Windows too. Stanford 2780 Skypark Drive, Torrance, CA 90505
Graphics takes full advantage of the latest P R E 5 E N T A T I O N .. A N A L Y 5 I 5 Tel: (310) 325-1339 FAX: (310) 325-1505

© 1992 SCIENTIFIC AMERICAN, INC

This content downloaded from

192.188.53.214 on Tue, 04 Jul 2023 15:50:28 +00:00
All use subject to https://about.jstor.org/terms