Artificial Intelligence: Genetic Programming

In the meantime, it is an appealing idea that Pollack J B 1990 Recursive distributed representations. Artificial
representation and learning techniques from both Intelligence 46(1–2): 77–106
symbolic processing models and connectionist net- Quillian M 1968 Semantic memory. In: Minsky M (ed.) Semantic
work models shall be brought together to tackle prob- Information Processing. MIT Press, Cambridge, MA, pp.
216–60
lems that neither type of model alone can apparently Quinlan R 1986 Inductive learning of decision trees. Machine
handle very well. Learning 1: 81–106
One such problem is the modeling of human Russell S J, Norvig P 1995 Artificial Intelligence: A Modern
cognition, which requires dealing with a variety of Approach. Prentice-Hall, Englewood Cliffs, NJ
cognitive capacities. Several researchers (e.g., Smol- Shavlik J W, Dietterich T G 1990 Reading in Machine Learning.
ensky 1988, Dreyfus and Dreyfus 1986) have con- Morgan Kaufmann, San Mateo, CA
sistently argued that cognition is multifaceted and Shastri L, Ajjanagadde V 1993 From simple associations to
better captured with a combination of symbolic and systematic reasoning: A connectionist representation of rules,
connectionist processes. Many methods and fram- variables and dynamic bindings using temporal synchrony.
eworks reviewed above share the belief that conn- Beha
ioral and Brain Sciences 16(3): 417–51
Smolensky P 1988 On the proper treatment of connectionism.
ectionist and symbolic methods can be usefully in- Beha
ioral and Brain Sciences 11(1): 1–23
tegrated, and such integration may lead to significant Sun R 1994 Integrating Rules and Connectionism for Robust
advances in our understanding of cognition. It thus Commonsence Reasoning. Wiley, New York
appears that hybridization is a theoretically sound Sun R, Alexandre F (eds.) 1997 Connectionist Symbolic In-
approach, in addition to being a practically expedient tegration. Erlbaum, Mahwah, NJ
approach. Sun R, Bookman L A (eds.) 1995 Computational Architectures
Integrating Neural and Symbolic Processes. Kluwer,
Boston
See also: Artificial and Natural Computation; Arti- Sun R, Peterson T 1998 Autonomous learning of sequential
ficial Intelligence: Connectionist and Symbolic Ap- tasks: Experiments and analyzes. IEEE Transactions on Neural
proaches; Artificial Intelligence in Cognitive Science; Networks 9(6): 1217–34
Artificial Neural Networks: Neurocomputation; Turing A 1950 Computing machinery and intelligence. Mind 59:
Connectionist Approaches; Connectionist Models of 433–60
Concept Learning; Connectionist Models of Lan- Waltz D, Feldman J A (eds.) 1988 Connectionist Models and Their
guage Processing; Intelligence: History of the Con- Implications. Ablex, Norwood, NJ
Zadeh L A 1988 Fuzzy logic. Computer 21(4): 83–93
cept; Knowledge Representation; Scientific Discovery,
Computational Models of R. Sun

Bibliography
Barnden J A, Pollack J B (eds.) 1991 Ad
ances in Connectionist
and Neural Computation Theory. Ablex Publishing Co., Artificial Intelligence: Genetic
Norwood, NJ
Davis E 1990 Representations of Commonsense Knowledge. Programming
Morgan Kaufmann, San Mateo, CA
Dreyfus H L, Dreyfus S E 1986 Mind O
er Machine. Free
Press, New York 1. Introduction
GilesC L,Gori M1998 Adapti
eProcessingof Sequences and Data
Structures. Springer, New York
The term genetic programming (GP) describes a
Lange T, Dyer M 1989 High-level inferencing in a connectionist research area within artificial intelligence (AI) that
network. Connection Science 1: 181–217 deals with the evolution of computer code. The term
McCarthy J L 1968 Programs with common sense. In: Minsky M evolution refers to an artificial process analogous to
(ed.) Semantic Information Processing. MIT Press, Cambridge, natural evolution of living organisms, but which has
MA been abstracted and stripped of most of its intricate
Rumelhart D E, McClelland J L, PDP Research Group 1986 details. The resultant algorithms then yield approxi-
Parallel Distributed Processing: Explorations in the Micr- mate solutions to problems in machine learning or
ostructures of Cognition. MIT Press, Cambridge, MA induce precise solutions in the form of grammatically
Michalski R S 1983 A theory and methodology of inductive correct (language) structures for the automatic pro-
learning. Artificial Intelligence 20: 111–61
gramming of computers.
Miikkulainen R, Dyer M 1991 Natural language processing with
modular PDP networks and distributed lexicons. Cogniti
e
Genetic programming is part of the growing set of
Science 15(3): 343–99 evolutionary algorithms which apply search principles
Newell A, Simon H 1976 Computer science as empirical inquiry: analogous to those of natural evolution in a variety of
Symbols and search. Communication of ACM 19: 113–26 different problem domains, notably parameter optimi-
Pearl J 1988 Probabilistic Reasoning in Intelligent Systems. zation. Evolutionary programming, evolutionary
Morgan Kaufmann, San Mateo, CA strategies, and genetic algorithms are three other

789
Artificial Intelligence: Genetic Programming

branches of the area of evolutionary algorithms which 2178940928 2178840928

mostly find applications as optimization techniques. 2953183257 Mutation 2953183257
All evolutionary algorithms follow Darwin’s prin- 2179465216 2179465216
ciple of differential natural selection. This principle 2177359880 2177359880
states that the following preconditions must be fulfilled 16777216 16777216
for evolution to occur via (natural) selection:
There are entities called individuals which form a
population. These entities can reproduce or can be
reproduced.
There is heredity in reproduction, that is to say that
individuals produce similar offspring. 2176543589 2176543589
In the course of reproduction there is variety which 2323488798 2323488798
affects the likelihood of survival and therefore of 2875642322 Crossover 2179465216
reproducibility of individuals.
2387907764 217359880
There are finite resources which cause the individ-
uals to compete. Due to overreproduction of individ- 16777216
uals not all can survive the struggle for existence.
Differential natural selection exerts a continuous Figure 2
pressure towards improved individuals. The primary operations of GP, mutation, and
The representation of programs, or generally struc- crossover are applied here to programs represented by
tures, in GP has a strong influence on the behavior and the sequences of instructions. The instructions are
efficiency of the resulting algorithm. As a consequence, coded as integer numbers which allows easy
many different approaches toward choosing repre- manipulation by access to these numbers
sentations have been adopted in GP. The resulting
principles have been applied even to other problem methods for selection, both deterministic and stoch-
domains such as design of electronic circuits or art and astic. Selection determines (a) which programs are
musical composition. allowed to survive (overproduction selection), and (b)
which programs are allowed to reproduce (mating
selection). Once a set of programs has been selected for
2. The Mechanisms Behind Genetic Programming further reproduction, the following operators are
applied: reproduction, mutation, and crossover.
Genetic programming works with a population of Reproduction simply copies an individual, mutation
programs that are executed or interpreted in order to varies the structure of an individual under control
judge their behavior. Usually, a scoring operation of a random number generator, and crossover mixes
called fitness measurement is applied to the outcome the structure of two (or more) programs to generate
of the behavior. For instance, the deviation between one or more new programs (see Fig. 2). Additional vari-
the quantitative output of a program and its target ation operators are applied in different applications.
value (defined through an error function) could be Most of these contain problem knowledge in the
used to judge the behavior of the program. This is form of heuristic search recipes which are adapted
straightforward if the function of the target program to the problem domain.
can be clearly defined. Results may also be defined as In this way, fitness advantages of individual pro-
side-effects of a program, such as consequences of the grams are exploited in a population to lead to better
physical behavior of a robot controlled by a genetically solutions. A key effort in GP is the definition of the
developed program. Sometimes, an explicit fitness fitness measure. Sometimes the fitness measure has to
measure is missing, for instance in a game situation, be iteratively improved in order for the evolved
and the results of the game (winning or losing) are solutions actually to perform the function they were
taken to be sufficient scoring for the program’s intended for. The entire process can be seen in close
strategy. The general approach is to apply a variety of analogy to breeding animals. The breeder has to select
programs to the same problem and to compare their those individuals from the population which carry the
performance relative to each other (see Fig. 1). targeted traits to a higher degree than others.
The outcomes of fitness measurement are used to Researchers (Turing 1950, Friedberg 1958, Samuel
select programs. There are a number of different 1959) suggested ideas similar to GP already in the
early days of AI, but did not get very far. So it was only
Input after other techniques in evolutionary algorithms had
Variation devive Loop Selection device been successfully developed that GP emerged. Earlier
Output
work concerned genetic algorithms (Holland 1975),
evolution strategies (Schwefel 1981), and evolutionary
Figure 1 programming (Fogel et al. 1966). These methods have
The variation selection loop of GP been applied successfully to a wide spectrum of

790
 Artificial Intelligence: Genetic Programming

problem domains, especially in optimization. How- schema theory of genetic algorithms (Goldberg 1989,
ever, it was unclear for a long time whether the Vose 1999) has been a primary target of knowledge
principles of evolution could be applied to computer transfer. In the meantime, a number of different
code, with all its dependencies and structural brittle- schema theorems have been formulated for GP.
ness. Negative results from early experiments seemed When analyzing search spaces of programs it was
to indicate that evolution of computer code was not realized that their size is many orders of magnitude
possible. Successes were all in the area of constraint larger than search spaces of combinatorial optimiz-
optimization (Michalewicz 1996), where methods were ation problems. A typical size for a program search
made available for how to deal with structural brittle- space might be 10"!!,!!!, as opposed to a typical search
ness. These methods found their way into program- space for a combinatorial optimization problem of the
ming and gave rise to the new field of GP (Koza 1992). order of 10"!!. Although this might be interpreted as
discouraging for search mechanisms, it was also
realized that the solution density in program spaces is,
3. Progress and State-of-the-art above a certain threshold, constant with changing
complexity (Langdon 1999). In other words, there are
In his seminal work of 1992, Koza established the field proportionally many more valid solutions in program
of GP by arguing convincingly that manipulation of spaces than in the spaces of combinatiorial optimiz-
symbolic tree structures is possible with evolutionary ation problems.
algorithms and that the resulting technique would
have a wide variety of applications. In subsequent
years, the field experienced both broadening and 4. Applications
deepening (Banzhaf et al. 1998). Many different
representations for GP were studied, among them The main application areas of GP are (from narrow to
other generic data structures such as sequences of wide) (Banzhaf et al. 1998): computer science, science,
instructions or directed graphs, as well as more exotic engineering, and art and entertainment. In computer
data structures such as stacks or neural networks. science, the development of algorithms has been a
Today, different approaches are considered as GP, focus of attention. By being able to manipulate
from the evolution of parse trees to the evolution of symbolic structures, genetic programming is one of the
arbitrary structures. The overarching principle is to few heuristic search methods for algorithms. Sorting
subject structures with variable complexity to forces of algorithms, caching algorithms, random number
evolution by applying mutation, crossover, and generators, and algorithms for automatic parallel-
fitness-based selection. The results must not necess- ization of code (Ryan 2000), to name a few, have been
arily be programs. studied. The spectrum of applications in computer
An ever-present difficulty with GP is that the science spans from the generation of proofs for
evolution of structures of variable complexity leads predicate calculus to the evolution of machine code for
often to large structures with considerable redun- accelerating function evaluation. The general tendency
dancy. Notably, variable complexity often leads to is to try to automate the design process for algorithms
inefficient and space-consuming code. It was subse- of different kinds.
quently recognized that the evolutionary forces
exerted a pressure toward more complex solutions, Typical applications in science are to modeling and
most of which could be removed after evolution pattern recognition. Modeling certain processes in
without doing any harm to the evolved solution. By physics and chemistry with the unconventional help of
drawing an analogy from biological evolution of evolutionary creativity supports research and under-
genomes, this phenomenon was called ‘intron growth,’ standing of the systems under study. Pattern rec-
or growth of ineffective code. Thought the explanation ognition is a key ability in molecular biology and other
for this phenomenon is not fully understood yet, it was branches of biology, as well as in science in general.
found that at least two different influences were at Here, GP has delivered first results that are competitive
work promoting the growth of complexity during if not better than human-generated results.
evolution. The most important one has to do with the In engineering, GP is used in competition or
protection effect of redundant code if subjected to the cooperation with other heuristic methods such as
action of crossover or mutation. Redundant code was neural networks or fuzzy systems. The general goal is
resistant to crossover and mutation and allowed its again to model processes such as production plants, or
carrier solution to survive better, compared to other to classify results of production. Control of man-made
individuals which did not possess the redundancy. apparatus is another area where GP has been used
Currently, many researchers are working to transfer successfully, with process control and robot control
results from research in genetic algorithms to genetic the primary applications.
programming. To achieve results in GP is generally In art and entertainment, GP is used to evolve
more difficult since it works with variable complexity realistic animation scenes and appealing visual graph-
and multiple fitness cases for fitness scoring. The ics. It also has been used to extract structural in-

791
Artificial Intelligence: Genetic Programming

formation from musical composition in order to model Bibliography

the process so that automatic composition of music
Banzhaf W, Nordin P, Keller R, Francone F 1998 Genetic
pieces becomes possible. Programming—An Introduction on the Automatic E
olution of
Many of these problems require a huge amount of Computer Programs and its Applications. Morgan Kaufmann,
computational power on the part of the GP systems. San Francisco, CA
Parallel evolution has hence been a key engineering Fogel L J, Owens A, Walsh M 1966 Artificial Intelligence
aspect of developments in GP. As a paradigm, genetic Through Simulated E
olution. Wiley, New York
programming is very well suited for a natural way of Friedberg R M 1958 A learning machine: part I. IBM Journal of
parallelization. With the advent of inexpensive parallel Research and De
elopment 2: 2–13
hard- and software (Sterling et al. 1999), a considerable Gilbert N G, Troitzsch K G 1999 Simulation for the Social
proliferation of results is expected from GP systems. Scientist. Open University Press, Buckingham, UK
Goldberg D E 1989 Genetic Algorithms in Search, Optimization
and Machine Learning. Addison Wesley, Reading, MA
Holland J H 1975 Adaptation in Natural and Artificial Systems:
an Introductory Analysis with Applications to Biology, Control,
and Artificial Intelligence. University of Michigan Press, Ann
5. Methodological Issues and Future Directions Arbor, MI
In recent years, some researchers have claimed human- Koza J R 1992 Genetic Programming: on the Programming
competitive results in the application of genetic pro- of Computers by Means of Natural Selection. MIT Press,
gramming to certain problems (Koza et al. 1999). Cambridge, MA
Koza J R, Andre D, Bennett F, Keane M 1999 Genetic
These claims are based on a comparison between the Programming III: Darwinian In
ention and Problem Sol
ing.
best known at present human solution to a problem Morgan Kaufmann, San Francisco, CA
and its GP counterpart. Usually, a large amount of Langdon W 1999 Boolean function fitness spaces. In: Poli R,
computational power had to be invested in order to Nordin P, Langdon W, Fogarty T (eds.) Proceedings
gain human-competitive results from genetic program- EuroGP’99. Springer, Berlin, pp. 1–14
ming runs. Due to the ability of the human mind Michalewicz Z 1996 Genetic AlgorithmsjData Structures l
quickly to grasp the recipes of problem solution an E
olution Programs. Springer, Berlin
artificial system has applied, the question remains Ryan C 2000 Automatic Re-engineering of Software Using
open whether GP solutions will stay better than human Genetic Programming. Kluwer, Boston, MA
solutions. Samuel A 1959 Some studies in machine learning using the game
of checkers. IBM Journal of Research and De
elopment 3:
Theory of genetic programming is presently greatly 210–29
underdeveloped and will need to progress quickly in Schwefel H P 1981 Numerical Optimization of Computer Models.
order to catch up with other evolutionary algorithm Wiley, Chichester, UK
paradigms. Most of the obstacles stem from the fact of Sterling T L, Salmon J, Becker D, Savarese D 1999 How to Build
variable complexity of solutions evolved in GP. a Beowulf: A Guide to the Implementation and Application of
Implementation of GP will benefit in the coming years PC Clusters. MIT Press, Cambridge, MA
from new approaches which include research from Turing A M 1950 Computing machinery and intelligence. Mind
developmental biology. Also, it will be necessary to 59: 433–60
learn to handle the redundancy forming pressures in Vose M D 1999 The Simple Genetic Algorithm: Foundations and
Theory. MIT Press, Cambridge, MA
the evolution of code.
Application of genetic programming will continue
W. Banzhaf
to broaden. Many applications focus on controlling
behavior of real or virtual agents. In this role, genetic
Copyright # 2001 Elsevier Science Ltd.
programming may contribute considerably to the
growing field of social and behavioral simulations. All rights reserved.
Genetic algorithms have already been found beneficial
in optimizing strategies of social agents (Gilbert and
Troitzsch 1999). With its ability to adjust the com- Artificial Intelligence in Cognitive Science
plexity of a strategy to the environment and to allow
competition between agents, GP is well positioned to Artificial intelligence (AI) and cognitive science are, at
play an important role in the study and simulation of the turn of the twenty-first century, two distinct
societies and their evolution. disciplines, with overlapping methodologies, but
rather different goals. AI is a branch of computer
science and is concerned with the construction and
See also: Adaptation, Fitness, and Evolution; Arti- deployment of intelligent agents as computer pro-
ficial Intelligence: Connectionist and Symbolic Ap- grams, and also with understanding the behavior of
proaches; Artificial Intelligence: Search; Evolution: these artifacts. The core scientific goal of AI is to
Optimization; Evolution: Self-organization Theory; understand basic principles of intelligent behavior that
Evolutionary Theory, Structure of; Intelligence: His- apply equally to animal and artificial systems (Russell
tory of the Concept; Natural Selection and Norvig 2000). Almost all of the work is math-

792

International Encyclopedia of the Social & Behavioral Sciences ISBN: 0-08-043076-7