## An Overview of Evolutionary Computation (1996)

Venue: | Chinese Journal of Advanced Software Research (Allerton |

Citations: | 14 - 10 self |

### BibTeX

@ARTICLE{Yao96anoverview,

author = {Xin Yao},

title = {An Overview of Evolutionary Computation},

journal = {Chinese Journal of Advanced Software Research (Allerton},

year = {1996},

volume = {3},

pages = {12--29}

}

### Years of Citing Articles

### OpenURL

### Abstract

This paper presents a brief overview of the field of evolutionary computation. Three major research areas of evolutionary computation will be discussed; evolutionary computation theory, evolutionary optimisation and evolutionary learning. The state-of-the-art and open issues in each area will be addressed. It is indicated that while evolutionary computation techniques have enjoyed great success in many engineering applications, the progress in theory has been rather slow. This paper also gives a brief introduction to parallel evolutionary algorithms. Two models of parallel evolutionary algorithms, the island model and the cellular model, are described. 1 Introduction The field of evolutionary computation has grown rapidly in recent years [1, 2, 3]. Engineers and scientists with quite different backgrounds have come together to tackle some of the most difficult problems using a very promising set of stochastic search algorithms --- evolutionary algorithms (EAs). There are several diffe...

### Citations

10959 |
Computers and Intractability: A Guide to the Theory of NP-Completeness
- Garey, Johnson
- 1979
(Show Context)
Citation Context ...xperimental result and needs further investigation. It is well-known that no polynomial approximate algorithm exists for the TSP which guarantees the performance unless P = NP , a very unlikely event =-=[37]-=-. This result does not contradict Fogel's experimental result because the theoretical result is about the worst case time complexity and the experimental result is about the estimated average case tim... |

7363 |
Genetic Algorithms
- Goldberg
- 1989
(Show Context)
Citation Context ...le some of the most difficult problems using a very promising set of stochastic search algorithms --- evolutionary algorithms (EAs). There are several different types of EAs; genetic algorithms (GAs) =-=[4, 5]-=-, evolutionary programming (EP) [6, 7] and evolution strategies (ESs) [8, 9]. Each type has numerous variants due to different parameter settings and implementations. The answer to the question which ... |

3738 |
Stochastic relaxation, Gibbs distributions, and the Bayesian restoration of images
- Geman, Geman
- 1984
(Show Context)
Citation Context ...s [5, 28]. It is not difficult to analyse the global convergence of these elitist EAs using Markov chains [29]. The techniques used are very similar to those used to show the global convergence of SA =-=[30, 31]-=-. For non-elitist EAs, the analysis of global convergence is not straightforward. It has been shown that the simple GA [5] (also called the canonical GA) without elitism cannot converge to global opti... |

3558 | Optimization by simulated annealing
- Kirkpatrick, Gelatt, et al.
- 1983
(Show Context)
Citation Context ...ion to be no worse than the current one to be acceptable. SA does not have such a requirement. It regards a worse solution to be acceptable with certain probability. The difference among classical SA =-=[10]-=-, fast SA [11], very fast SA [12], and a new SA [13] is mainly due to the difference in their perturbations, i.e., methods of generating the next solution. EAs can be regarded as a population-based ve... |

1970 |
Genetic algorithms + data structures = evolution programs, third edn
- Michalewicz
- 1996
(Show Context)
Citation Context ...all problems [45]. The answer to the question whether an EA is efficient is problem dependent. Most problems attacked by evolutionary optimisation are unconstrained optimisation problems. Michalewicz =-=[46, 47]-=- has achieved some very good results with his EAs on unconstrained optimisation problems. Constraint handling is still one of the unsolved problems in evolutionary optimisation. Simply adding a fixed ... |

1176 | Handbook of Genetic Algorithms
- DAVIS
- 1991
(Show Context)
Citation Context ...us a good framework to investigate various crossover operators. For the binary encoding scheme, two-point crossover and uniform crossover have been empirically shown to be effective for many problems =-=[49]-=-, especially for numerical optimisation problems. Some combinatorial optimisation problems, such as the TSP and various scheduling problems, require certain kinds of order-based crossover operators [5... |

865 |
An Analysis of the Behavior of a Class of Genetic Adaptive Systems
- Jong
- 1975
(Show Context)
Citation Context ... are those which always copy the best individual in the current generation to the next generation. Such elitist EAs include EP for numerical optimisation [7, 23], (�� + )-ESs [9, 27], and elitist =-=GAs [5, 28]-=-. It is not difficult to analyse the global convergence of these elitist EAs using Markov chains [29]. The techniques used are very similar to those used to show the global convergence of SA [30, 31].... |

662 |
Evolutionsstrategie: optimierung technischer systeme nach prinzipien der biologischen evolution
- Rechenberg
- 1973
(Show Context)
Citation Context ...d nonelitist. Elitist EAs are those which always copy the best individual in the current generation to the next generation. Such elitist EAs include EP for numerical optimisation [7, 23], (�� + )-=-=ESs [9, 27]-=-, and elitist GAs [5, 28]. It is not difficult to analyse the global convergence of these elitist EAs using Markov chains [29]. The techniques used are very similar to those used to show the global co... |

646 | Intelligence through simulated evolution. Forty years of evolutionary programming - Fogel - 1999 |

605 | Tabu search
- Glover, Laguna
- 1998
(Show Context)
Citation Context ...his direction. 2.5 Analysis of Neighbourhood Structures and Landscapes Analysis of neighbourhood structures and landscapes has been carried out for many algorithms like SA [16, 17, 18, 19, 40] and TS =-=[41]-=-. Jones [15] recently proposed a formal model of landscapes for EAs. The model make it possible to analyse various EAs with different search operators, especially crossover operators. It provides a co... |

511 |
Numerical optimization of computer models
- Schwefel
- 1981
(Show Context)
Citation Context ...tic search algorithms --- evolutionary algorithms (EAs). There are several different types of EAs; genetic algorithms (GAs) [4, 5], evolutionary programming (EP) [6, 7] and evolution strategies (ESs) =-=[8, 9]-=-. Each type has numerous variants due to different parameter settings and implementations. The answer to the question which EA is the best is problem dependent. There is no universally best algorithm ... |

474 |
Evolution and Optimum Seeking
- Schwefel
- 1995
(Show Context)
Citation Context ...tic search algorithms --- evolutionary algorithms (EAs). There are several different types of EAs; genetic algorithms (GAs) [4, 5], evolutionary programming (EP) [6, 7] and evolution strategies (ESs) =-=[8, 9]-=-. Each type has numerous variants due to different parameter settings and implementations. The answer to the question which EA is the best is problem dependent. There is no universally best algorithm ... |

440 |
Co-evolving Parasites Improve Simulated Evolution as an Optimization Procedure
- Hillis
- 1992
(Show Context)
Citation Context ...eal-world benchmark problems. 4.2.1 Co-evolutionary Learning Co-evolutionary learning has two different forms. The first one indicates the situation where two populations are evolved at the same time =-=[68]-=-. The fitness of an individual in one population depends on the individuals in another population. There is no crossover or other information exchange between two populations. This can be regarded as ... |

333 | The GENITOR algorithm and selection pressure: Why rank-based allocation of reproductive trials is best
- Whitley
- 1989
(Show Context)
Citation Context ...s not depend on the raw fitness of an individual. It depends on the rank of an individual in the population. There are several variants of rank-based selection, such as Baker's [52, 53] and Whitley's =-=[54]-=- linear ranking and Yao's [51] nonlinear ranking. Difference rank-based selection mechanisms have different selection pressures. These selection pressures are the same across generations because they ... |

252 |
The evolution of strategies in the iterated prisoner’s dilemma
- Axelrod
- 1987
(Show Context)
Citation Context .... The second form of co-evolution is at the individual level. There is only one population involved. The fitness of an individual in the population depends on other individuals in the same population =-=[69, 62, 63]-=-. For example, the same strategy for playing an iterated prisoner's dilemma game may get quite different fitness values depending on what other strategies are in the population. Both forms of co-evolu... |

245 | No Free Lunch Theorems for Search
- Wolpert, Macready
- 1995
(Show Context)
Citation Context ...oo many to cite them all here). However, there are a few papers which contain some negative results about EAs [42, 43, 44]. This is not surprise because no algorithm can be efficient for all problems =-=[45]-=-. The answer to the question whether an EA is efficient is problem dependent. Most problems attacked by evolutionary optimisation are unconstrained optimisation problems. Michalewicz [46, 47] has achi... |

234 | A genetic algorithm tutorial
- Whitley
- 1994
(Show Context)
Citation Context ...e first and the last fixed position in the schema. It facilitates our discussion if we view one generation of a GA as having two steps; the first is selection and the second is crossover and mutation =-=[20]-=-: Current Population selection =) Intermediate Population crossover;mutation =) New Population 2.1.1 Impact of Selection on Schema Propagation Let f(H) be the average fitness of the samples in the pop... |

212 |
Adaptive selection methods for genetic algorithms
- Baker
- 1985
(Show Context)
Citation Context ...ank-based selection does not depend on the raw fitness of an individual. It depends on the rank of an individual in the population. There are several variants of rank-based selection, such as Baker's =-=[52, 53]-=- and Whitley's [54] linear ranking and Yao's [51] nonlinear ranking. Difference rank-based selection mechanisms have different selection pressures. These selection pressures are the same across genera... |

188 |
The parallel genetic algorithm as function optimizer
- Mühlenbein, Schomisch, et al.
- 1991
(Show Context)
Citation Context ...has better local search ability. SA has been used as such a local searcher for GAs [55]. The proposed genetic annealing algorithm [55] compared favourably with the GA or SA alone. M��uhlenbein et =-=al. [56]-=- used a simple hill-climbing algorithm as the local searcher for his parallel GA and also achieved very good experimental results. Every search algorithm, except for uniform random search, introduces ... |

181 | Very fast simulated re-annealing
- Ingber
- 1989
(Show Context)
Citation Context ...nt one to be acceptable. SA does not have such a requirement. It regards a worse solution to be acceptable with certain probability. The difference among classical SA [10], fast SA [11], very fast SA =-=[12]-=-, and a new SA [13] is mainly due to the difference in their perturbations, i.e., methods of generating the next solution. EAs can be regarded as a population-based version of generate-and-test search... |

168 |
Evolutionary Algorithms, Fitness Landscapes and Search
- Jones
- 1995
(Show Context)
Citation Context ...esearchers in terms of GA deceptive problems [5]. In recent years, analysis of neighbourhood structures and landscapes has attracted increasing attentions from both the evolutionary computation field =-=[15]-=- and other fields like SA [16, 17, 18, 19]. 2.1 The Schema Theorem The schema theorem was first proposed by Holland [4] to explain how GAs work by propagating similarity templates in populations. A sc... |

154 | A review of evolutionary artificial neural networks
- Yao
- 1993
(Show Context)
Citation Context ...lete CS. The whole population includes a number of competing CSs. 4.2 Population-Based Machine Learning Population-based machine learning includes CSs, evolutionary artificial neural networks (EANNs) =-=[61]-=-, and other evolutionary [62, 63] or non-evolutionary [64] systems. EANNs can be considered as a combination of artificial neural networks (ANNs) and EAs. EAs have been introduced into ANNs at three d... |

117 |
Incorporating problem specific knowledge into genetic algorithms
- Grefenstette
- 1987
(Show Context)
Citation Context .... In practice, a good way to tailor the generate-and-test search to the problem we are interested in is to incorporate problem specific heuristic knowledge into search operators and selection schemes =-=[14]-=-. The rest of this paper is organised as follows: Section 2 discusses some theoretical issues in evolutionary computation, including the classical schema theorem, convergence of EAs, neighbourhood sea... |

98 | Genetic algorithms and very fast simulated reannealing: A comparison
- Ingber, Rosen
- 1992
(Show Context)
Citation Context ...he results are mixed. Most papers have reported very positive results obtained from EAs (too many to cite them all here). However, there are a few papers which contain some negative results about EAs =-=[42, 43, 44]-=-. This is not surprise because no algorithm can be efficient for all problems [45]. The answer to the question whether an EA is efficient is problem dependent. Most problems attacked by evolutionary o... |

95 |
Biases in the crossover landscape
- Eshelman, Caruana, et al.
- 1989
(Show Context)
Citation Context ...olutionary algorithms with different operators also have different biases. There have been some discussions on the biases introduced by various crossover operators in the genetic algorithm literature =-=[57]-=-. It is such biases that make an algorithm very efficient for one class of problems (i.e., the biases lead to the efficient search of a class of fitness landscapes) but not for others. Since we do not... |

76 |
Fast simulated annealing
- Szu, Hartley
(Show Context)
Citation Context ...orse than the current one to be acceptable. SA does not have such a requirement. It regards a worse solution to be acceptable with certain probability. The difference among classical SA [10], fast SA =-=[11]-=-, very fast SA [12], and a new SA [13] is mainly due to the difference in their perturbations, i.e., methods of generating the next solution. EAs can be regarded as a population-based version of gener... |

76 |
Markov Chains: Theory and Applications
- Isaacson, Madsen
- 1975
(Show Context)
Citation Context ...elitist EAs include EP for numerical optimisation [7, 23], (�� + )-ESs [9, 27], and elitist GAs [5, 28]. It is not difficult to analyse the global convergence of these elitist EAs using Markov cha=-=ins [29]-=-. The techniques used are very similar to those used to show the global convergence of SA [30, 31]. For non-elitist EAs, the analysis of global convergence is not straightforward. It has been shown th... |

62 |
Evolving Artificial Intelligence
- Fogel
- 1992
(Show Context)
Citation Context ...time n, f X is the fitness of X , S is the whole search (solution) space, and S is the set of global optima. It has been shown that such global convergence can be established under certain conditions =-=[21, 22, 23, 24, 25, 26]-=-. As indicated by Rudolph [26], EAs can be classified into elitist and nonelitist. Elitist EAs are those which always copy the best individual in the current generation to the next generation. Such el... |

61 | Travelling salesman and sequence scheduling: quality solutions using genetic edge recombination, in: Handbook of Genetic Algorithms
- Whitley, Starkweather, et al.
- 1991
(Show Context)
Citation Context ...9], especially for numerical optimisation problems. Some combinatorial optimisation problems, such as the TSP and various scheduling problems, require certain kinds of order-based crossover operators =-=[5, 50, 51]-=-. The crossover operator for a combinatorial optimisation problem often needs to be designed specifically for the problem. Such design may not be easy because combinatorial problems usually have some ... |

54 | The time complexity of maximum matching by simulated annealing
- SASAK, HAJEK
- 1988
(Show Context)
Citation Context ...design of new EAs. From the point of view of algorithm design, we are more interested in the computational complexity of EAs for a particular problem. A result similar to that for simulated annealing =-=[32]-=- would be very valuable. 2.3 Computational Complexity Computational complexity is one of the most important issues in the analysis of algorithms. To our best knowledge, the only paper on this topic is... |

53 | An Experimental Study of N-Person Iterated Prisoner's Dilemma Games
- Darwen, Yao
- 1995
(Show Context)
Citation Context ... includes a number of competing CSs. 4.2 Population-Based Machine Learning Population-based machine learning includes CSs, evolutionary artificial neural networks (EANNs) [61], and other evolutionary =-=[62, 63]-=- or non-evolutionary [64] systems. EANNs can be considered as a combination of artificial neural networks (ANNs) and EAs. EAs have been introduced into ANNs at three different levels: the evolution of... |

51 | An Empirical Comparison of Seven Iterative and Evolutionary Function Optimization Heuristics,” CMU-CS95-193. Available via. http://www.cs.cmu.edu/~baluja
- Baluja
- 1995
(Show Context)
Citation Context ...he results are mixed. Most papers have reported very positive results obtained from EAs (too many to cite them all here). However, there are a few papers which contain some negative results about EAs =-=[42, 43, 44]-=-. This is not surprise because no algorithm can be efficient for all problems [45]. The answer to the question whether an EA is efficient is problem dependent. Most problems attacked by evolutionary o... |

49 |
A simulated annealing like convergence theory for the simple genetic algorithm
- Davis, Principe
- 1991
(Show Context)
Citation Context ...time n, f X is the fitness of X , S is the whole search (solution) space, and S is the set of global optima. It has been shown that such global convergence can be established under certain conditions =-=[21, 22, 23, 24, 25, 26]-=-. As indicated by Rudolph [26], EAs can be classified into elitist and nonelitist. Elitist EAs are those which always copy the best individual in the current generation to the next generation. Such el... |

48 |
Global convergence of genetic algorithms: A Markov chain analysis
- Eiben, Aarts, et al.
- 1991
(Show Context)
Citation Context ...time n, f X is the fitness of X , S is the whole search (solution) space, and S is the set of global optima. It has been shown that such global convergence can be established under certain conditions =-=[21, 22, 23, 24, 25, 26]-=-. As indicated by Rudolph [26], EAs can be classified into elitist and nonelitist. Elitist EAs are those which always copy the best individual in the current generation to the next generation. Such el... |

47 |
System Identification through Simulated Evolution: A Machine Learning Approach to Modeling
- Fogel
- 1991
(Show Context)
Citation Context ... using a very promising set of stochastic search algorithms --- evolutionary algorithms (EAs). There are several different types of EAs; genetic algorithms (GAs) [4, 5], evolutionary programming (EP) =-=[6, 7]-=- and evolution strategies (ESs) [8, 9]. Each type has numerous variants due to different parameter settings and implementations. The answer to the question which EA is the best is problem dependent. T... |

46 | On evolving robust strategies for iterated prisoner’s dilemma
- Darwen, Yao
(Show Context)
Citation Context ... includes a number of competing CSs. 4.2 Population-Based Machine Learning Population-based machine learning includes CSs, evolutionary artificial neural networks (EANNs) [61], and other evolutionary =-=[62, 63]-=- or non-evolutionary [64] systems. EANNs can be considered as a combination of artificial neural networks (ANNs) and EAs. EAs have been introduced into ANNs at three different levels: the evolution of... |

42 | Applying evolutionary programming to selected travelling salesman problems
- Fogel
- 1993
(Show Context)
Citation Context ...s difficult because it must be done for a particular problem. It does not make sense when we discuss the computational complexity of an EA without indicating which problem the EA is applied to. Fogel =-=[34, 35, 36]-=- has carried out some empirical studies on the computational complexity of his EP in solving selected travelling salesman problems (TSPs). He found that the time used to arrive at a good solution for ... |

39 |
An evolutionary approach to the travelling salesman problem
- Fogel
- 1988
(Show Context)
Citation Context ...s difficult because it must be done for a particular problem. It does not make sense when we discuss the computational complexity of an EA without indicating which problem the EA is applied to. Fogel =-=[34, 35, 36]-=- has carried out some empirical studies on the computational complexity of his EP in solving selected travelling salesman problems (TSPs). He found that the time used to arrive at a good solution for ... |

33 |
Efficient simulated annealing on fractal energy landscapes. Algorithmica 6, 367–418. ECCC ISSN 1433-8092 http://www.eccc.uni-trier.de/eccc ftp://ftp.eccc.uni-trier.de/pub/eccc ftpmail@ftp.eccc.uni-trier.de, subject ’help eccc
- Sorkin
- 1991
(Show Context)
Citation Context ...ptive problems [5]. In recent years, analysis of neighbourhood structures and landscapes has attracted increasing attentions from both the evolutionary computation field [15] and other fields like SA =-=[16, 17, 18, 19]-=-. 2.1 The Schema Theorem The schema theorem was first proposed by Holland [4] to explain how GAs work by propagating similarity templates in populations. A schema is a similarity template describing a... |

33 | Convergence of non-elitist strategies
- Rudolph
(Show Context)
Citation Context |

26 |
The only challenging problems are deceptive: Global search by solving order-1 hyperplanes
- Das, Whitley
- 1991
(Show Context)
Citation Context ...ms and GA-Hardness Although few researchers in the evolutionary computation field have been working on computational complexity, there are many people working on GA deceptive problems and GA-hardness =-=[5, 38, 39]-=-. The motivation for such studies is to identify hard problems for GAs and develop new algorithms to combat them. Nearly all GA deceptive problems investigated so far are artificial and created by hum... |

25 |
A markov chain analysis of a genetic algorithm
- Suzuki
- 1993
(Show Context)
Citation Context |

25 | Optimizing an Arbitrary Function is Hard for the Genetic Algorithm
- Hart, Belew
- 1991
(Show Context)
Citation Context ...able. 2.3 Computational Complexity Computational complexity is one of the most important issues in the analysis of algorithms. To our best knowledge, the only paper on this topic is by Hart and Belew =-=[33]-=-. The analysis of computational complexity for EAs is difficult because it must be done for a particular problem. It does not make sense when we discuss the computational complexity of an EA without i... |

24 | A population-based learning algorithm which learns both architectures and weights of neural networks
- Liu, Yao
- 1996
(Show Context)
Citation Context ...e different levels: the evolution of connection weights, architectures, and learning rules. At present, most work on EANNs concentrates on the evolution of architectures, i.e., connectivities of ANNs =-=[65, 66, 67]-=-. Very good results have been achieved for some artificial and real-world benchmark problems. 4.2.1 Co-evolutionary Learning Co-evolutionary learning has two different forms. The first one indicates t... |

23 |
Convergence properties of canonical genetic algorithms
- Rudolph
- 1994
(Show Context)
Citation Context |

21 | Simulated annealing with extended neighbourhood
- Yao
- 1991
(Show Context)
Citation Context ...ptive problems [5]. In recent years, analysis of neighbourhood structures and landscapes has attracted increasing attentions from both the evolutionary computation field [15] and other fields like SA =-=[16, 17, 18, 19]-=-. 2.1 The Schema Theorem The schema theorem was first proposed by Holland [4] to explain how GAs work by propagating similarity templates in populations. A schema is a similarity template describing a... |

19 | An empirical study of genetic operators in genetic algorithms
- Yao
- 1993
(Show Context)
Citation Context ...9], especially for numerical optimisation problems. Some combinatorial optimisation problems, such as the TSP and various scheduling problems, require certain kinds of order-based crossover operators =-=[5, 50, 51]-=-. The crossover operator for a combinatorial optimisation problem often needs to be designed specifically for the problem. Such design may not be easy because combinatorial problems usually have some ... |

17 | A preliminary study on designing artificial neural networks using coevolution
- Yao, Shi
- 1995
(Show Context)
Citation Context ...e different levels: the evolution of connection weights, architectures, and learning rules. At present, most work on EANNs concentrates on the evolution of architectures, i.e., connectivities of ANNs =-=[65, 66, 67]-=-. Very good results have been achieved for some artificial and real-world benchmark problems. 4.2.1 Co-evolutionary Learning Co-evolutionary learning has two different forms. The first one indicates t... |

16 |
A new simulated annealing algorithm
- Yao
- 1995
(Show Context)
Citation Context ...able. SA does not have such a requirement. It regards a worse solution to be acceptable with certain probability. The difference among classical SA [10], fast SA [11], very fast SA [12], and a new SA =-=[13]-=- is mainly due to the difference in their perturbations, i.e., methods of generating the next solution. EAs can be regarded as a population-based version of generate-and-test search. They use search o... |

16 |
An Analysis of the Effects of Selection in Genetic Algorithms
- Baker
- 1989
(Show Context)
Citation Context ...ank-based selection does not depend on the raw fitness of an individual. It depends on the rank of an individual in the population. There are several variants of rank-based selection, such as Baker's =-=[52, 53]-=- and Whitley's [54] linear ranking and Yao's [51] nonlinear ranking. Difference rank-based selection mechanisms have different selection pressures. These selection pressures are the same across genera... |