You are on page 1of 4

Generating Biomorphs with an Aestheti

Immune Systemy
Dennis L. Chao


Stephanie Forrest

Department of Computer S ien e

University of New Mexi o
Albuquerque, NM 87131 USA

fdl hao,forrestg

Abstra t
We des ribe an intera tive sear h algorithm inspired by
the immune system. The algorithm learns what parts
of the sear h spa e are not useful to help a user explore large parameter spa es e iently. The algorithm
is apable of nding onsensus solutions for parties with
di erent sele tion riteria. A simple implementation of
the algorithm applied to sele ting Biomorphs (Dawkins
1986) is presented.
Introdu tion

In this paper we illustrate how an immune systeminspired ltering algorithm des ribed in (Chao & Forrest 2002) an be applied to the ltering and generation
of art. When the immune system is exposed to a novel
pathogen, it learns to re ognize it in what is alled the
primary response, and it remembers the pattern so that
subsequent infe tions an be eliminated more e iently
in the se ondary response. Analogously, an arti ial immune system an be viewed as \prote ting" users from
undesirable data. We propose onstru ting an arti ial
immune system that learns what kinds of data patterns
are una eptable and then prevents exposure to similar data in the future. This approa h s ales to multiple
users by using the superset of the undesirable data of
many users to lter data. The data that survive this
ensoring is the set that ould satisfy everyone. These
solutions are useful for situations in whi h one solution
must satisfy several people at on e, su h as musi that
is broad ast to an audien e or artwork that is displayed
in publi spa es.
We demonstrate the prin iples of this idea by applying it to the generation of simple gures known as
Biomorphs (Dawkins 1986). The system is apable of
learning whi h regions of Biomorph parameter spa e produ e images that are displeasing to a parti ular user and
an generate diverse andidate solutions of high quality.
Furthermore, the solutions produ ed by ombining the
pro les of a small group of users are likely to be pleasing
to all of the group members.
To appear in Arti ial Life VIII: Pro eedings of the
Eighth Inte rnational Conferen e on the Simulation and
Synthesis of Living Systems

An Aestheti Immune System

An aestheti immune system is a lter between users and

a \stream" of art and eliminates undesired art before it
an rea h the user. Like the adaptive immune system,
it \learns" from experien e. If a user labels a work of
art as \bad," the system will remember this and prevent
this work from being shown to the user in the future.
It would not be pra ti al for the user to evaluate every
item manually, so the algorithm must generalize. If the
user dislikes a ertain item, it is likely that the user will
dislike similar items as well.
Aestheti spa e is a framework for grouping \similar"
works of art. In this spa e, the distan e between the
lo ations of two works of art is proportional to their
similarity. Thus, similar works are lose to ea h other
while dissimilar works are far apart. For example, oil
paintings of the Renaissan e ould o upy points in one
part of the spa e, the paintings of the Enlightenment
era might not be too far away, while the work of Abstra t Expressionists would be quite distant from both
of these. Clearly this spa e is impossible to onstru t
a urately due to the subje tive nature of similarity in
the realm of aestheti s. However, when generating art
algorithmi ally, we will assume that for many systems
similar parameters produ e aestheti ally similar images.
If a user reje ts one work, he or she will probably
dislike most of its neighbors in aestheti spa e. Therefore
the system reates a negative dete tor that overs a small
portion of spa e surrounding ea h reje ted work. The
dete tors will behave like lympho ytes in the immune
system, re ognizing anything that is too similar to their
pre-programmed targets. If a work is not reje ted by any
of the dete tors, it is allowed to \survive." Therefore,
the user should never be presented a new work that is
similar to one expli itly reje ted in the past.
The works of art that an pass through the system are
not similar to previously reje ted ones, but they are not
guaranteed to be \good." If the user reje ts one of these,
the system will form a new dete tor for it, whi h will prevent similar \bad" works from being seen in the future.
The untrained system with no dete tors gradually onstru ts a pro le of the user's tastes as it sees examples

of art that the user reje ts. Eventually, the a umulated

dete tors will form a model of the user's preferen es and
shield the user from a broad range of undesirable art.
To use an aestheti immune system to help a user
explore a wide variety of potentially good solutions, a
random number generator an supply a stream of random andidate solutions to it. The system would lter
out andidates that are likely to be bad, and when fully
trained (if this is possible), it is a perpetual generator
of novel but satisfa tory art. Note that it does not attempt to nd optimal solutions, so other te hniques may
be used to re ne solutions presented to the user.
An important onsequen e of negative dete tion is
that dete tors an be independently generated by users
with di erent judgment riteria. The areas that are not
ex luded by the superset of their negative dete tors an
de ne the regions of \ onsensus" solutions, or solutions
that are satisfa tory to everyone. Many users an generate \bad art" dete tors independently, and when the
dete tor sets of many users are ombined, the system will
generate works that none of the users will dislike. One
an think of ea h user's dete tor set a ting as a sieve,
blo king the passage of art that is disagreeable to that
user. The solutions that pass through all the sieves are
the onsensus solutions. This method ould potentially
satisfy the aestheti demands of multiple users e iently
and a urately, even if the users are spe ifying their preferen es asyn hronously and in rementally. The quality
of the onsensus solutions depends on the domain and
the heterogeneity of the users. If the users have on i ting interests, onsensus solutions might not exist.
Related work

Evolutionary art (evoart) usually relies on humans to

provide the feedba k needed by the omputer to guide
the \evolution" of a work of omputer generated art.
Typi ally, the user iteratively re nes a work by sele ting a subset of favorite works out of a small set, whi h
are variants or ombinations of the user's favorites from
the previous time step. After many iterations, the quality of the works an improve. This pro ess of \evolving" art is sometimes alled aestheti sele tion. Dawkins
was the rst to implement evolutionary art on a omputer (Dawkins 1986). In his system, the user explores
the spa e of images alled \Biomorphs" in the iterative
manner des ribed above.
The aestheti immune system approa h has several advantages over the evolutionary approa h. First, evoart
systems usually require a user to ompare many members of a population at on e to pi k a favorite, while
an immunologi ally-based system does not. For some
systems, su h as the generation of musi al phrases, it
is di ult to evaluate many individuals at on e (Nelson
1993). Se ond, the evolutionary pro ess onverges to single works of art, while the immunologi al approa h eval-

uates the whole of parameter spa e and an be a sour e

of perpetual novelty. Third, in order to produ e innovative works, evolutionary systems often ombine works of
art (Sims 1991). The programmer is fa ed with the di ult task of de ning operators to mix the \genotypes" of
works in a manner su h that the o spring's \phenotype"
exhibits desirable traits of its parents. In some ases,
the results are unpredi table and displeasing, suggesting that more omplex ombination rules are required.
A trained aestheti immune system an generate a very
large range of aestheti ally pleasing art without the use
of a di ult-to-de ne art ombination operator. This
may be ru ial in ollaborative evoart, in whi h many
users ooperate to generate works of art. Simply ombining the favorite works of di erent users might yield
unsatisfa tory results. If a fan of Salvador Dal's surrealist anvases were to meet a fan of Norman Ro kwell's
sentimental and more realisti style, a painting that ombined the two painters' tenden ies would probably be
unsatisfying to both. A better solution might be to introdu e both to something ompletely di erent, like the
paintings of Paul Klee. This is the sort of solution an
aestheti immune system would be apable of proposing.
Appli ation to Biomorphs

We applied the immunologi al framework to the exploration of Dawkins' Biomorphs (Dawkins 1986). A
Biomorph is a re ursively drawn gure de ned by a ninedigit \genotype." The digits, all but one of whi h are
allowed to take the values between -9 and 9 in lusively,
are parameters that a e t the appearan e of the nal
image, or \phenotype." The nal digit of the genotype,
whi h de nes the level of re ursion used in drawing the
Biomorph, is restri ted to values between 2 and 9. The
following se tions des ribe how to implement an aestheti immune system to help users nd aestheti ally
pleasing regions of the Biomorph genotype spa e.

De ning the aestheti distan e

The rst step is to de ne the distan e between any two

solutions. Fortunately, small hanges in Biomorph parameters (the genotype) usually result in small phenotypi hanges, so we were able to use genotype spa e as
a proxy for the true aestheti spa e. That is, Biomorphs
with similar genotypes are assumed to look similar, while
the appearan e of Biomorphs with very di erent genotypes is assumed un orrelated. For onvenien e, we de ne the distan e in genotype spa e to be the maximum
arithmeti di eren e between any two orresponding digits of the two genotypes, whi h is the Minkowski metri
1 )1=1 , where a and b are the
bi j
L1 (a; b) = [
i (jai
strings being ompared and ai and bi are the ith element of a and b. For example, the distan e between the
two genotypes (0,0,0,0,1,8,1,1,2) and (7,3,6,2,1,1,3,3,2) is
7 be ause that is the maximum di eren e between the
orresponding digits of the two strings (at both positions

1 and 6). Other distan e measures are possible, su h as

the sum of the di eren es between orresponding positions in the genotype.
For some domains, one would not want to generate
new works, su h as with a olle tion of already existing
work. In su h ases, the immunologi al approa h ould
be used to prevent the retrieval of entries that are too
similar to previously reje ted works. There is a large
body of resear h on the problem of quantifying similarity between data, in luding ontent-based lassi ation,
metadata-based lassi ation, and ollaborative ltering
te hniques.

De ning the dete tors

A dete tor should be large enough so that the user does
not need to de ne \too many" of them but not so large
that they eliminate as many good solutions as bad ones.
We will de ne a dete tor of radius r to over all genotypes within distan e r from the enter of the dete tor.
The neighborhood size of a dete tor is the number of
distin t lo ations in spa e that it overs. The dete tor
neighborhood should over a \large enough" portion of
the total parameter spa e su h that a reasonable number
of dete tors ould over a large portion of the total genotype spa e. It may be fruitful to augment the immunologi al approa h with pattern lassi ation te hniques to
more e iently over parameter spa e.
For the Biomorph user experiment, des ribed below,
we used large dete tors in order to minimize the amount
of time it took the subje ts to over a signi ant portion
of spa e. We hose a dete tor size su h that the average
user ould reate enough dete tors to ll about half of
Biomorph parameter spa e in under ten minutes. The
large dete tor size a e ts the results quantitatively, but
hopefully not qualitatively. It is probably impossible to
guarantee that no good genotype will be ensored by a
dete tor, but we would like to minimize this possibility.
If the transition between high and low quality solutions
is not too abrupt, there ould be regions of expendable
intermediate-quality solutions between them in parameter spa e. If the user is instru ted to not reje t solutions
of low but a eptable quality, the dete tors are less likely
to interfere with good solutions.
We set the dete tor radius to be 6 for all parameters
ex ept for the last one, whi h we set to 2. Small hanges
in the last parameter greatly a e t the appearan e of
Biomorphs, so a smaller radius is appropriate. Figure 1
shows a random Biomorph and various others that fall
within its dete tor radius. Based on this and several
other examples, we on luded that Biomorphs that are
within the range of a single dete tor are qualitatively
similar and assume that if a genotype yields an undesirable phenotype, then most genotypes within its dete tor
are likely to be undesirable as well.

Figure 1: A sample from the range of Biomorphs overed

by a single dete tor. The boxed gure in the upper-left
is the dete tor's enter, whi h is (3,6,6,-5,3,-2,-6,-5,8).

Seven volunteers parti ipated in a user study to determine the feasibility of the aestheti immune system. All
used a program to reate their own Biomorph dete tor sets and were retested several days later to determine the e e tiveness of system's dete tors. To reate a
Biomorph dete tor set, a omputer program displayed
a series of randomly generated Biomorphs. For ea h
Biomorph shown, the subje t was given the option to
reje t it, reating a dete tor based on its genotype, or to
a ept it, whi h does not hange the state of the dete tor
set. The dete tors were put to use as they were reated
to lter out new Biomorphs similar to the reje ted ones.
Thus, the aestheti quality of images was expe ted to improve during a session be ause the in reasing number of
dete tors ould prevent more regions of \bad" genotypes
from being displayed. The subje ts were asked to rate
Biomorphs until about 50% of the parameter spa e was
overed by dete tors. To test the e e tiveness of the
dete tors, the subje ts were asked to evaluate a series
of 150 Biomorphs from three sets: random Biomorphs,
Biomorphs that survived ensoring by the subje t's own
dete tors, and Biomorphs that survived the ensoring by
the superset of many subje ts' dete tors. The number of
reje ted Biomorphs from ea h set was re orded, and we
assume user satisfa tion to be inversely orrelated with
the number of reje tions.
This test was performed using all seven subje ts and
the superset of all subje ts' dete tors to determine the
e a y of the onsensus solutions. It was repeated using
only three of the subje ts and their dete tors to determine the e e ts of group size on the quality of onsensus
solutions. The onsensus solutions shown in the latter
test were thus ensored using only three sets of dete tors
instead of seven.












Table 1: The user study results. The entries represent

the per entage of Biomorph images reje ted by a user
for a parti ular dete tor set. Ea h olumn represents the
data from one subje t. The dete tor sets, indi ated by
the labels \none", \self" and \all", are no dete tors, the
subje t's own dete tors, and the superset of all subje ts'
dete tors respe tively. The results from testing all seven
users are summarized in a), and the the results from the
subset of three are shown in b).

The results of the experiments are summarized in Table 1. The system appeared to learn the Biomorph preferen es of individual users to a statisti ally signi ant
degree. For the testing involving all seven users, the reje tion rate for the random Biomorphs was higher than
that for Biomorphs that were ltered using a subje t's
own dete tors (> 95% signi an e using the Wil oxon
signed-rank test). For single users, the system learly
redu ed the number of una eptable solutions, thus improving the average quality of solutions presented to an
individual. We believe that the system's performan e
ould be improved dramati ally for motivated users who
are willing to invest more time to build their personal
dete tor sets (e.g. for musi , movies, or books that they
like.) Using smaller dete tors and overing a larger portion of the parameter spa e would likely in rease the
a ura y of user preferen es. More importantly, a more
a urate aestheti distan e measure needs to be de ned.
Combining user preferen es worked well for the set of
three users but not for all seven. For the testing involving the seven users, the di eren e between reje tion
rates of random Biomorphs and those ensored by the superset of all seven subje ts' dete tors is not statisti ally
signi ant. For the testing involving only three of the
seven subje ts, the reje tion rates for the group-sele ted
Biomorphs were approximately the same or lower than
the ones sele ted using their own dete tors. The onsensus solutions for three users appeared to be of equal
or greater quality than the solutions generated for the
individuals alone, while the solutions for all seven users
appeared to be of low quality. The results suggest that
it is hard to satisfy a large number of users and that
onsensus solutions may not even exist for groups with

members with di erent preferen es. The number of users

that an aestheti immune system an a ommodate is
likely dependent on the parti ular problem domain and
the similarity of the users. If the users have on i ting
preferen es or if the problem domain is highly subje tive, then onsensus solutions would be hard to nd.
However, if user preferen es are orthogonal to one another, then a large range of good onsensus solutions
ould be dis overed using this method. Ina ura ies in
the user pro les an also lower the average quality of
solutions found. If the pro les for all users were perfe t
and omplete, solutions that are not ensored by any
pro le would satisfy everyone. However, the Biomorph
pro les were approximate and in omplete, so ombining
pro les did not always work.
Con lusions

We have presented an algorithm for identifying regions of

interest in a parameter spa e using the subje tive judgments of multiple individuals. It remains to be seen if
the onsensus solutions are more likely to be novel and
innovative or bland and ino ensive. It would be easy to
retro t many existing evoart algorithms to use the immunologi al approa h. We believe that this will reveal
interesting parts of the parameter spa es that were not
seen before due to the ine ien y of mutational sear h
and the lo al aestheti tness optima that an stall evolution. The sear h for these regions ould be a elerated
by having multiple users reate a ommon pool of dete tors. If the set of ooperating users is large and representative of the tastes of the general publi , the nal
result might be a system that generates art that many
an appre iate.
A knowledgments

We thank the parti ipants in the user study. The authors gratefully a knowledge the support of the NSF
(grants CDA-9503064 and ANIR-9986555), the ONR
(grant N00014-99-1-0417), DARPA (grant AGR F3060200-2-0584), the Intel Corporation, and the Santa Fe Institute.
Referen es

Chao, D. L., and Forrest, S. 2002. Information immune systems. In Pro eedings of the First International Conferen e on Arti ial Immune Systems. (in
Dawkins, R. 1986. The Blind Wat hmaker. Harlow,
UK: Longman S ienti and Te hni al.
Nelson, G. L. 1993. Sonomorphs: An appli ation of
geneti algorithms to the growth and development of
musi al organisms. In Pro eedings of the Fourth Biennial Art & Te hnology Symposium, 155{169.
Sims, K. 1991. Arti ial evolution for omputer graphi s. Computer Graphi s 25(4):319{328.