Welcome to Scribd. Sign in or start your free trial to enjoy unlimited e-books, audiobooks & documents.Find out more
Standard view
Full view
of .
Look up keyword
Like this
0 of .
Results for:
No results containing your search query
P. 1
Evolutionary Dynamics

Evolutionary Dynamics



|Views: 209|Likes:

More info:

Published by: William Tanksley, Jr on Mar 07, 2007
Copyright:Attribution Non-commercial


Read on Scribd mobile: iPhone, iPad and Android.
download as PDF, TXT or read online from Scribd
See more
See less





Sufficient Conditions for Coarse-GrainingEvolutionary Dynamics
Keki BurjorjeeDEMO Lab,Computer Science Department,Brandeis University, Waltham, MA 02454.kekib@cs.brandeis.edu
Previous theoretical results in the Evolutionary Computation liter-ature only permit analyses of evolutionary dynamics in the
— i.e. over a single generation — or in the asymptote of time.There are currently no theoretical results that permit a
anal-ysis of any non-trivial aspect of evolutionary dynamics in the
short term 
, i.e. over a small number of generations. In the absence of suchanalyses we believe that accurate theories of evolutionary adaptationwill continue to evade discovery. We describe a technique called coarse-graining which has been widely used in other scientific disciplines tostudy the emergent phenomena of complex systems. This techniqueis a promising approach towards the formulation of more principledtheories of evolutionary adaptation because, if successfully applied, itpermits a principled analysis of evolutionary dynamics across multiplegenerations. We develop a simple yet powerful abstract framework forstudying the dynamics of an infinite population evolutionary algorithm(IPEA). Using this framework we show that the short term dynamicsof an IPEA can be coarse-grained if it satisfies certain abstract condi-tions. We then use this result to argue that the dynamics of an infinitepopulation
algorithm with uniform crossover and fitness pro-portional selection can be coarse-grained for at least a small numberof generations, provided that the initial population belongs to a par-ticular class of distributions (which includes the uniform distribution),and the fitness function satisfies a relatively
Simple Genetic Algorithms (GAs) have successfully been used to adaptsolutions for a wide range of search problems. One of the most impor-
tant open questions, one might argue
most important question, inGA research today is the question of how these algorithms performadaptation. Complete theories of how the iterated effect of selectionand variation on an initially random population drives adaptation havebeen scarce in the thirty odd years since GAs were first proposed. Weknow of just one work in which a theory of adaptation has been com-pletely laid out — the seminal work of Holland [9], in which a theoryof adaptation, that later came to called the Building Block Hypothesis[8,11] was proposed. The Building Block Hypothesis, though well-formulated, is not nec-essarily accurate. It has been sharply criticized for lacking theoretical justification and experimental results have been published that drawits veracity into question. On the theoretical side, for example, Wrightet. al. state in [20], “The various claims about GAs that are tradi-tionally made under the name of the
building block hypothesis
have, todate, no basis in theory, and in some cases, are simply incoherent.”.On the experimental side Syswerda has reported in[16] that uniformcrossover often outperforms one-point and two-point crossover on thefitness functions that he has studied. Summing up Syswerda’s resultsFogel remarks [7,p.140] “Generally, uniform crossover yielded betterperformance than two-point crossover, which in turn yielded betterperformance than one-point crossover”. These results contradict thebuilding block hypothesis because uniform crossover is extremely dis-ruptive of short schemata whereas one and two-point crossover arecertainly more likely to conserve short schemata and combine theirdefining bits in children produced during recombination.
The Absence of Principled Theories of Adaptation
A very general way of understanding the absence of principled the-ories of adaptation in the Evolutionary Computation literature is tonote firstly that the set of GAs comprise a class of 
complex systems
,secondly that adaptation is an
emergent phenomenon 
of many (butnot all) of the systems in this class, and finally that it is in generalvery difficult to formulate accurate theories of how or why particularemergent phenomena arise (or do not arise) from the dynamics of thecomplex systems in some class[2].The state of any system in a class of complex systems is typicallymodeled as a tuple of 
state variables. The dynamics of the systemsin this class is modeled by a system of 
parameterized coupled dif-ference or differential equations. A set of parameter values determinesa particular dynamical system, and when these values are ‘plugged-in’to the equations they describe how the state variables of that systemchange over time. One (dead-end) approach to understanding howsome target phenomenon arises through the dynamics of a class of 
complex systems is to attempt to solve the set of equations, i.e. toattempt to obtain a closed form formula which, given some parametervalues that determine some system and the state of that system attime step
= 0 (the initial condition), gives the state of the systemat some time-step
. In principle, one can then attempt to under-stand the target phenomenon by studying the closed-form solution.The equations of a complex system however are always non-linear, andtypically unsolvable, so this approach has a low likelihood of success.Another approach is to attempt to glean an understanding of the targetphenomenon by
numerically simulating 
the dynamics of a well chosensubset of the systems in the class under a well chosen subset of ini-tial conditions. This approach becomes infeasible as
, the number of state variables, becomes large.The dynamics of an EA can be modeled by a system of coupled non-linear difference equations called an infinite population model (IPEAs).The time-dependent state variables in such a system is the expectedfrequencies of individual genomes
. The simulation of one generationof an IPEA with
state variables has time complexity
). Aninfinite population model of a GA (IPGA) with bitstring genomes of length
= 2
state variables. Hence, the time complexity for a‘naive’ numeric simulation of an IPGA is
). (See [19,p.36] for adescription of how the Fast Walsh Transform can be used to bring thisbound down to
).) Even when the Fast Walsh Transform is used,computation time still increases exponentially with
. Vose reportedin 1999 that computational concerns force numeric simulation to belimited to cases where
20.Given this discission, we are now in a position to give a more specificreason for the absence of principled theories of adaptation in Evolu-tionary Computation. Current results in the field only allow one toanalyze aspects of IPEA dynamics at the asymptote of time (e.g.[19])or in the
immediate term 
, i.e. over the course of a single generation(e.g.[10,14]). For IPEAs with large genome sizes and non-trivial fitness functions, there are currently no theoretical results that permita
analysis of 
non-trivial aspect of IPEA dynamics inthe
short term 
(i.e. over a small number of generations). We believehowever that a principled analysis of the short term is necessary inorder to formulate accurate theories of adaptation. In the absence of results that permit such analysis we believe that accurate theories of adaptation will continue to evade discovery.
Evolutionary Biologists use the word genome to refer to the total hereditary informa-tion that is encoded in the DNA of an individual, and we will as well. This same conceptis called a genotype in the evolutionary computation literature. Unfortunately this usagecreates a terminological inconsistency. The word genotype is used in evolutionary biologyto refer to a different concept, one that is similar to what is called a schema in evolutionarycomputation

Activity (2)

You've already reviewed this. Edit your review.
1 thousand reads
1 hundred reads

You're Reading a Free Preview

/*********** DO NOT ALTER ANYTHING BELOW THIS LINE ! ************/ var s_code=s.t();if(s_code)document.write(s_code)//-->