Goldberg, D. E., 1989a, Genetic algorithms and Walsh functions: Part II, deception and its analysis, Complex Syst. 3:153--171.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....Strategy (Shaefer [20] does alter the representation during the course of the search, but not in the manner suggested by Holland, nor in a way which is amenable to this analysis. Walsh function analysis is also sometimes used for post mortem analysis of why a genetic algorithm fails (Goldberg, [10]) Goldberg [8] however, suggested the following two principles for good representations: The Principle of Meaningful Building Blocks: The user should select a [representation] so that short, low order schemata are relevant to the underlying problem and relatively unrelated to schemata over ....

....two chromosomes as required (albeit with low probability) and it should be apparent that they always respect schemata. 7 Deception Deception, like most work on genetic algorithms, has only hitherto been considered in the context of classical schemata, and has been rigourously defined by Goldberg [10]. If, however, more general formae are considered, then it becomes necessary to consider deception in terms of the formae under consideration. Of course, Holland [15] advocated using inversion with one point crossover. The aim of this was to bring co adapted sets of alleles closer together on ....

D. E. Goldberg, Genetic Algorithms and Walsh Functions: Part II, Deception and Its Analysis, in Complex Systems 3:2 1989.

....in the case of selection and recombination we have f #e# x (t) P x (t) j x (t) 37) f #e# x (t) 1 and j x (t) t)P # x m (t) 38) The corresponding e#ective selection coe#cient is s # e# = 1 f(t) 1 . 39) This e#ective fitness is intuitively more similar to that of [112, 113, 114, 115] in that it takes into account only the destructive e#ect of crossover. Nevertheless, it captures a rather natural division into terms that lead to a multiplicative renormalization of reproductive fitness (destruction terms) and those that lead to an additive renormalization (creation terms) In ....

Goldberg, D. E. Genetic algorithms and Walsh functions: Part II. Deception and its analysis. Complex Systems, 3:153--171, 1989.

....operators they are most appropriate. Common measurements of problem di#culty are: Correlation analysis, polynomial decomposition and Walsh coe#cients, and . schemata analysis. These measurements of problem di#culty are widely used in the EA literature for measuring problem di#culty [15, 16, 17, 18, 5]. For an overview see [19] Their specific properties are briefly discussed in the following paragraphs. Correlation analysis is based on the assumption that the high and low quality solutions are grouped together and neighboring individuals have similar fitnesses. Problems are easy if the ....

D. E. Goldberg. Genetic algorithms and Walsh functions: Part II, deception and its analysis. Complex Systems, 3:153--171, 1989.

....complete basis set that satisfy some common closure properties that are satis ed by most of the common basis that we often deal with. However, in the following discussion we choose to work with Walsh basis functions because of its existing connections to the eld of genetic algorithms [8, 9, 14, 16, 17, 18, 26, 45, 58, 61]. Walsh basis is functionally complete over the space of all boolean strings. In other words it can represent any function that can be de ned over the space of boolean strings. The following discussion o ers a brief overview of Walsh representation. Walsh functions [5, 76] are orthogonal ....

....Laplace, and other transformations, Walsh functions are often used to represent a problem solving task in a convenient form. Application of Walsh transformation (WT) in understanding Genetic Algorithms was rst noted by Bethke [7] Further investigation of this approach can be found elsewhere [14, 16, 17, 26, 50, 61, 74, 75]. Traditionally, the Walsh functions are used for representing real valued functions of binary variables. However, they can be easily extended to higher cardinality representation, as shown elsewhere [58] Although the main arguments of the following discussion can be extended for higher ....

D. E. Goldberg. Genetic algorithms and Walsh functions: Part II, deception and its analysis. Complex Systems, 3(2):153-171, 1989. (Also TCGA Report 89001).

....a function and points out how the general problem of function induction relates to this representation in particular. 2. 3 The Multidimensional Fourier Transform and the Nonlinearity of Functions In this section, we recapitulate the Multidimensional Fourier Transform (MFT) or the Walsh transform [13, 14, 6, 19, 35, 43], which is a useful tool for detailed study of the functions we will be considering as well as the GCTs themselves. We use the MFT to study real valued functions de ned on X. Let F be the set of all such functions. F forms a ( Q n k i ) dimensional vector space over R. The MFT of a function f ....

D. Goldberg. Genetic algorithms and Walsh functions: part II. Complex Systems, 3:153-171, 1989.

....illustrate that a genetic algorithm converges to many distant points and that those points are only partially consistent with the low order schema information. 2 A Walsh Analysis of Satisfiability Problems A method for studying the epistasis in a binary function is to use Walsh analysis [5, 6, 9]. All binary functions can be represented as a weighted sum of Walsh Copyright c fl1998, Springer Verlag. http: www.springer.de comp lncs index.html functions denoted by j ; where 0 j 2 Gamma 1 with each Walsh function being j : B f Gamma1; 1g. The real valued weights are called ....

D. Goldberg. Genetic algorithms and walsh functions: Part II, deception and its analysis. Complex Systems, 3:153--171, 1989.

....linear contribution to the evaluation function associated with bit position 0, while w 0101 measures the nonlinear interaction between the bits in positions 0 and 2. This nonlinearity is often considered to be an important feature in determining problem difficulty for stochastic search algorithms [3, 4, 12]. 2 Summary Statistics for Problem Instances Walsh analysis can be used to compute summary statistics for fitness distributions of discrete optimization problems. Note that the fitness distribution is the distribution formed by evaluating all possible inputs to a problem. So, for a problem ....

D. Goldberg. Genetic Algorithms and Walsh Functions: Part II, Deception and its Analysis. Complex Systems, 3:153--171, 1989.

....1: e(0000) 28 e(0100) 20 e(1000) 12 e(1100) 4 e(0001) 26 e(0101) 18 e(1001) 10 e(1101) 2 e(0010) 24 e(0110) 16 e(1010) 8 e(1110) 0 e(0011) 22 e(0111) 14 e(1011) 6 e(1111) 30 The function is posed as a maximization problem. This function is not fully deceptive [4] [9] since ( 1) 0) and ( 11) 00) where ( is the average evaluation of all strings in hyperplane . However, it does have a significant amount of deception. The following function eb(i) represents evaluation B constructed from Function 1. eb(0000) e(0000) 28 eb(1000) e(0000) ....

D. Goldberg, (1989) Genetic algorithms and walsh functions: Part II, deception and its analysis. Complex Systems 3:153-171.

....to be solved provide misleading information. In this chapter, we will present an extension of the pcBHS model to cater for these two cases byintroducing redundancy and competition. 5.1 Introduction 5.1. 1 The challenge of deceptiveness Deception has been discussed rigorously in GA community[24,15,60,61,14,47, 25, 52] in the last decade. Briefl , deceptive problems contain deceptive attractors which mislead the algorithm to search for sub optima. Figure 5.1 shows a typical fully deceptive function for simple GAs on a maximization problem 1 . As shown in the figure, the global optimum and the suboptimum are ....

D.E. Goldberg. Genetic algorithms and walsh functions: Part ii. deception and its analysis. Complex systems, 3(2):153--171, 1989.

....cant and neglect its contribution. Fourier bases and their close relatives Walsh bases are frequently used to study the behavior of genetic algorithms. Walsh bases [4] were rst used by Bethke [6] for analyzing genetic algorithms. Further investigation of this approach can be found elsewhere [9, 11, 12, 14, 33, 34, 35, 43, 44]. 4.2 Function induction from data and Fourier basis Function induction from data plays an important role in adaptation, machine learning, and non enumerative black box optimization. In function induction, the goal is to learn a function f : X Y from the data set = f(x (1) y (1) x ....

D. E. Goldberg. Genetic algorithms and Walsh functions: Part II, deception and its analysis. Complex Systems, 3(2):153-171, 1989. (Also TCGA Report 89001).

....the GA evaluates a single organism string, it is, in effect, simultaneously calculating the average fitness of all the organism s substrings. Holland dubbed the GA s capacity to perform these simultaneous tasks implicit parallelism. For accessible introductions to schema theory, see [11] 9] and [10]. 3 The slot machine has nothing to do with this part its only role is in solving the problem. It plays no part in representing or posing the problem. The objective function is some black box that has nothing at all to do with the slot machine 4 One small difference between this model and ....

David E. Goldberg. Genetic algorithms and walsh functions: Part ii, deception and its analysis. Complex Systems, 3:153--171, 1989. 64

....of evaluation function F(x, # H t,T , # GA t,T ) This hypothesis, in fact, touches the core issue of what makes a problem di#cult for a GA. Several criteria, such as isolation, deception, multimodality [14] have been suggested as measures of di#culty level of a problem for a GA. See [10, 11, 13, 15, 20, 44] for more information. The function f(x, # H t,T , # GA t,T ) has a high degree of isolation where fruitful attractors are closely surrounded by points with low fitness value. The isolation level of the FEF F(x, # H t,T , # GA t,T )is low, while its level of deception (fruitless ....

D. E. Goldberg. Genetic algorithms and Walsh functions: Part II, deception and its analysis. Complex Systems, 3:153--171, 1989.

....complete basis set that satisfy some common closure properties that are satis ed by most of the common basis that we often deal with. However, in the following discussion we choose to work with Walsh basis functions because of its existing connections to the eld of genetic algorithms [8, 9, 14, 16, 17, 18, 26, 43, 54, 57]. Walsh basis is functionally complete over the space of all boolean strings and equivalent to other choices of basis functions in this space. The following discussion o ers a brief overview of Walsh representation. Walsh functions [5] are orthogonal functions that found applications in many ....

....Fourier, Laplace, and other transformations, Walsh functions are often used to represent the representation in a convenient form. Application of Walsh transformation (WT) in understanding Genetic Algorithms was rst noted by Bethke [7] Further investigation of this approach can be found elsewhere [14, 16, 17, 26, 48, 57]. Traditionally, the Walsh functions are used for representing real valued functions of binary variables. However, they can be easily extended to higher cardinality representation, as shown elsewhere [54] Although the main arguments of the following discussion can be extended for higher ....

D. E. Goldberg. Genetic algorithms and Walsh functions: Part II, deception and its analysis. Complex Systems, 3(2):153-171, 1989. (Also TCGA Report 89001).

....of the original version of mGA [19] deserve the credit for first realizing the importance of detecting appropriate relations among the members of the search space. Another interesting aspect of their work was the class of problems they wanted to solve. The class of bounded deceptive problems [15] captures the essence of order k delineability developed in SEARCH. They took the right steps in both designing algorithms and identifying a class of problems that can be solved efficiently when the representation is fixed. However, the mGA had many problems. Some of them are listed in the ....

D. E. Goldberg. Genetic Algorithms and Walsh Functions: Part II, Deception and Its Analysis. Complex Systems, 3(2):153--171, 1989. (Also TCGA Report 89001).

....cant and neglect its contribution. Fourier bases and their close relatives Walsh bases are frequently used to study the behavior of genetic algorithms. Walsh bases [4] were rst used by Bethke [6] for analyzing genetic algorithms. Further investigation of this approach can be found elsewhere [9, 11, 12, 14, 33, 34, 35, 43, 44]. 4.2 Function induction from data and Fourier basis Function induction from data plays an important role in adaptation, machine learning, and nonenumerative black box optimization. In function induction, the goal is to learn a function f : X Y 7 Protein feature mRNA codon 1 100 1 ....

D. E. Goldberg. Genetic algorithms and Walsh functions: Part II, deception and its analysis. Complex Systems, 3(2):153-171, 1989. (Also TCGA Report 89001).

....is to assume that a solution is made up of a number of building blocks. If these building blocks can be discovered independently and combined afterwards, we get a tractable problem. A difficult instance of this class can be created by using the parameterized set of fully deceptive trap functions [3]. A fully deceptive trap (sub)function of order k has value [8] f (x ) ae k if u(x ) k k Gamma u(x ) Gamma 1 otherwise where u(x ) is a function that counts the number of 1 bits in x . The global optimum of this function is the string consisting of k 1 bits resulting in the maximal ....

D.E. Goldberg. Genetic algorithms and walsh functions: Part II, deception and its analysis. Complex Systems, 3:153--171, 1989.

....Strategy (Shaefer [20] does alter the representation during the course of the search, but not in the manner suggested by Holland, nor in a way which is amenable to this analysis. Walsh function analysis is also sometimes used for postmortem analysis of why a genetic algorithm fails (Goldberg, [10]) Goldberg [8] however, suggested the following two principles for good representations: The Principle of Meaningful Building Blocks: The user should select a [representation] so that short, low order schemata are relevant to the underlying problem and relatively unrelated to schemata over ....

....two chromosomes as required (albeit with low probability) and it should be apparent that they always respect schemata. 7 Deception Deception, like most work on genetic algorithms, has only hitherto been considered in the context of classical schemata, and has been rigourously defined by Goldberg [10]. If, however, more general formae are considered, then it becomes necessary to consider deception in terms of the formae under consideration. 6 Of course, Holland [15] advocated using inversion with one point crossover. The aim of this was to bring co adapted sets of alleles closer together on ....

D. E. Goldberg, Genetic Algorithms and Walsh Functions: Part II, Deception and Its Analysis, in Complex Systems 3:2 1989.

....of the parameters, and GA hardness analysis. For GA, our main concern here, these analyses are based on dioeerent approaches : ffl Proofs of convergence based on Markov chain modeling [6, 3, 1, 20] ffl Deceptive functions analysis, based on Schema analysis and Holland s original theory [14, 8, 9, 11], which characterizes the eOEciency of a GA, and allows to shed light on iGA hardj functions. ffl Some rather new approaches are based on an explicit modelization of a GA as a dynamical system [16, 22] Deception has been intuitively related to the biological notion of epistasis [5] which can be ....

....easily nd it. On the contrary, if the intersection of these building blocks is a secondary optimum, the population will preferably converge onto it, missing the global one. In this situation the GA will be considered to have failed 1 and f will be called deceptive. More formally, Goldberg ( 8] [9]) dened the static deception : The selection results in an expected greater mean tness for the set of individuals selected for reproduction, than for the preceding population. But this mean value will be changed by the application of genetic operators. It follows that the GA can be considered as ....

[Article contains additional citation context not shown here]

David E. Goldberg. Genetic algorithms and walsh functions: Part ii, deception and its analysis. In Complex Systems, volume 3, pages 153171. 1989.

....positions 0 and 2, and so on. Therefore Equation 1 says that any function over L bit space can be represented as a weighted sum of all possible 2 L bit interaction functions j . This nonlinearity is an important feature in determining problem difficulty for genetic algorithms [Goldberg, 1989a, Goldberg, 1989b, Reeves and Wright, 1995] The 2 L Walsh coefficients can be computed by a Walsh transform: w j = 1 2 L 2 L Gamma1 X x=0 f(x) j (x) 2) The calculation of Walsh coefficients can be thought of in terms of matrix multiplication. Let f be a column vector of 2 L elements where ....

....statistics for large MAXSAT problems and use this to look for statistical indicators of a phase transition. The polynomial time Walsh analysis of embedded landscapes also means that we can exactly compute schema averages upto a fixed order in polynomial time [Rana et al. 1998] Goldberg, 1989a] Goldberg, 1989b] Knowing exact schema averages and exact summary statistics for any particular problem instance actually provides a significant amount of information. On the other hand, despite having all this information, the results of Hastad [Hastad, 1997] indicates that in the general case, no search ....

Goldberg, D. (1989b). Genetic algorithms and walsh functions: Part ii, deception and its analysis. Complex Systems, 3:153--171.

.... (the so called Gray coding debate; Caruana Schaffer, 1988; though see Culberson, 1996, who shows that this argument is in fact futile without additional information) and discussion about whether a real valued coding might be more natural in this domain (with, e.g. Davis, 1991b, arguing for, and Goldberg, 1990, against) Applications like the travelling salesman problem, with naturally non orthogonal representations, have also highlighted this issue (e.g. Grefenstette et al. 1985) This led to practical invest23 igations of various aspects of representation. Shaefer (1987) developed the adaptive ....

....in order that the fitness of binary chromosomes can then be calculated by decoding them. However, such shoe horning may make much of the structure of the search problem unavailable to the algorithm in terms of heritable allele patterns (see for example the discussion of meaningful alphabets in Goldberg, 1990). For problems in which candidate solutions are more complicated objects (e.g. the travelling sales rep problem) a direct binary encoding may be unnatural or even infeasible. A particular case is when the natural variables of the problem are not orthogonal, in that the valid settings of one ....

[Article contains additional citation context not shown here]

D. E. Goldberg, 1989b. Genetic algorithms and Walsh functions: Part II, deception and its analysis. Complex Systems, 3:153--171.

....but not in others. The problem is that we cannot tell the difference simply from the SS, and we need the auxiliary information about the magnitude and sign of the effects to get a clearer picture of the difficulty of a particular problem. In [1] we showed that Goldberg s 3 bit deceptive function [7] is characterized by a large (negative) 3 gene interaction, and by conditions that imply combinations of interactions must have a net effect greater than the main effects. This can be extended to longer strings, and in general it can be shown that a class of hard deceptive functions may be ....

....Those NK landscapes that gave a high value on the variance metric could not be confirmed as difficult because the aliasing prevented the proper identification of the effects. Since we have shown in [1] that the ED approach is equivalent to the Walsh transform analysis popularized by Goldberg [7, 9], we should point out that this conclusion applies equally to methods based on Walsh functions as to the approach of Davidor which has been given prominence in this paper. 4 Further Implications The implications for epistasis measurement are clear, but this also has some relevance to ideas of ....

D.E.Goldberg (1989) Genetic algorithms and Walsh functions: part II, deception and its analysis. Complex Systems, 3, 153-171.

.... most notably Lawrence Davis (e.g. Davis, 1991) and perhaps the sharpest controversy has centred upon whether it is appropriate to use binary representations whether traditional or Gray coded (Caruana Schaffer, 1988) for problems in which the parameters under consideration are real (Goldberg, 1990c) In the Evolutionstrategie school, a more relaxed attitude towards alternative operators and representations appears always to have been taken (Baeck et al. 1991) It is worth pointing out that it has always been accepted even by the mainstream of the American school that there are certain ....

....(1989) Radcliffe (1991a) and Vose (1991) and is detailed below in section 3. 2.3. Representations and Operators Vose Liepins (1991) have pointed out that the difference between the simplest problems for genetic search and those normally considered to be the hardest (fully deceptive problems, Goldberg, 1990a, 1990b) is no more than a change of representation. Vose (1991) has also introduced the notion of a global schema, which is one immune to sampling error by virtue of being fitter than average in an arbitrary population if it is fitter than average in any specific population. It is clear that the ....

Goldberg, 1990b. David E. Goldberg. Genetic algorithms and walsh functions: Part II, deception and its analysis. Complex Systems, 3:153--171, 1990.

....of the parameters, GA easy or GA difficulty analysis. For GA, our main concern here, these analyses are based on different approaches : ffl proof of convergence based on Markov chain modeling [12, 8, 1, 45] ffl deceptive functions analysis, based on Schema analysis and Holland s original theory [23, 18, 19, 21], which characterizes the efficiency of a GA, and sheds light on GA difficult functions. ffl some rather new approaches are based on an explicit modeling of a GA as a dynamic system [28, 32, 57] It has to be noted first that in the modeling of GA as dynamic systems, some fractal features have ....

....to obtain even very preliminary clues that allow the parameters of a stochastic optimization algorithm like GA to be tuned, in order to perform an efficient optimization on such functions. 1.2. 2 GA Deception analysis of Holder functions This analysis is based on Goldberg s deception analysis [18, 19], which uses a decomposition of the function to be optimized, f , on Walsh polynomials. This decomposition allows the definition of a new function f 0 , which reflects the behaviour of the GA, and which represents the expected fitness value that can be reached from the point x : f 0 (x) ....

[Article contains additional citation context not shown here]

D. E. Goldberg. Genetic Algorithms and Walsh functions : Part II, deception and its analysis. TCGA Report No. 89001, University of Alabama, Tuscaloosa, US, 1989.

....In terms of schema: a problem is deceptive when it has building blocks which lead away from an optimal solution. The formal analysis of deception originates with Bethke s [Bet80] introduction of Walsh 3 INTRODUCTION TO COMPUTATIONAL COMPLEXITY 6 transforms, an approach which Goldberg [Gol89a, Gol89b] explains admirably and Bridges and Goldberg [BG91] extend. More recently, Holland [Hol89] introduced hyperplane transforms, which also make a rigorous analysis of deception possible. See Whitely [Whi91] for an excellent survey of deception. Liepins and Vose [LV91] have proven that some degree of ....

David E. Goldberg. Genetic algorithms and walsh functions: Part II, deception and its analysis. Complex Systems, 3:153--171, 1989.

.... of convergence based on Markov chain modeling : for example, Davis [7] has established a mutation probability decreasing scheme that ensures the theoretical convergence of the canonical algorithm, ffl deceptive functions analysis, based on schema analysis and Holland s original theory [16] 11] [12], 13] which characterizes the efficiency of a GA, and allows shedding light on GA difficult functions. Deception has been intuitively related to the biological notion of epistasis [6] which can be understood as a sort of nonlinearity degree. Deception depends on : the parameter setting of ....

....800 1000 Fitness Integer representation of the chromosomes Fig. 2. Onemax function on 10 bits : the sampling of a Holder function, with h = 0 (the abscissa is the usual integer representation of a binary string) III. Deception Analysis Our approach is based on Goldberg s deception analysis [11] [12], which uses a decomposition of the function to be optimized, f , on Walsh polynomials. This decomposition allows defining a new function f 0 , which can be understood as a sort of statistic preference given by the GA to the points of the search space during the search. This function f 0 is ....

[Article contains additional citation context not shown here]

D. E. Goldberg. Genetic Algorithms and Walsh functions : Part II, Deception and its Analysis. TCGA Report No. 89001, University of Alabama, Tuscaloosa, US, 1989.

.... based on Markov chain modeling : for example, Davis [7] has shown that a very low decreasing of the mutation probability p m along the generations insures the theoretical convergence onto a global optimum, ffl deceptive functions analysis, based on Schema analysis and Holland s original theory [15, 10, 11, 12], which characterizes the efficiency of a GA, and allows to shed light on GA difficult functions. Deceptivity has been intuitively related to the biological notion of epistasis [6] which can be understood as a sort of non linearity degree. Deceptivity depends on : the parameter setting of ....

....it is possible to go a little further and to find an a posteriori validation rule not only for the sampling precision, but also for the other parameters of the method. This is what we present in the next sections. 4 Deceptivity Analysis Our approach is based on Goldberg s deceptivity analysis [10, 11], which uses a decomposition of the function to be optimized, f , on Walsh polynomials. This decomposition allows to define a new function f 0 , which can be understood as a sort of preference given by the GA to the points of the search space during the search. This function f 0 is in some ....

[Article contains additional citation context not shown here]

D. E. Goldberg. Genetic Algorithms and Walsh functions : Part II, deception and its analysis. TCGA Report No. 89001, University of Alabama, Tuscaloosa, US, 1989.

.... better solutions There have been a lot of works dealing with this problem, both theoretical and empirical studies, initiated by a seminal paper by Davidor [1] His epistasis variance have been shown to have a strong mathematical foundation, based on Walsh functions analysis (see [2 4] and also [5, 6]) but still the correlation between problem hardness and epistasis is not straightforward. As it was argued in [3] even if it may provide some guidance for this matter, no definitive conclusion can be drawn. This is also confirmed by recent works from Rochet and Venturini [7] showing that one ....

David E. Goldberg. Genetic algorithms and Walsh functions: Part II, deception and its analysis. Complex Systems, 3:153--171, 1989.

....and negatively to an individual s fitness. Negative building blocks have negative fitness contributions that decrease an individual s fitness. However, a negative building block may be part of a larger, positive building block. This situation is similar to the deceptive problems described in (Goldberg 1989) where building blocks that are discouraged encouraged early in a run may may not lead to the optimum individual. Initial experiments showed 0 1000 2000 3000 4000 5000 6000 1 2 3 4 Average number of generations Negative level GA 32x8.FRR2 (nclen 5) GA 32x8.FRR2 (nclen 10) Figure 3: Number ....

Goldberg, D. E. 1989. Genetic algorithms and walsh functions: Part II, deception and its analysis. Complex Systems 3:153--171.

....or not useful schema relationships exist. Computing the exact average fitnesses of low order schemata requires exponential time for most large problems. However, we can use Walsh coefficients to efficiently compute low order schema averages. Functions ff and fi are defined on a schema, h, as per Goldberg (1989a) ff(h) i] ae 0 if h[i] 1 if h[i] 0 or 1 fi(h) i] ae 0 if h[i] or 0 1 if h[i] 1 For example, consider h = 01 : Because schema h is a subset of strings, its fitness is the average fitness of all strings that belong in that subset. Rather than enumerating the subset of strings, the fitness of ....

Goldberg, D. 1989b. Genetic algorithms and walsh functions: Part ii, deception and its analysis. Complex Systems 3:153--171.

....of the parameters, and GA hardness analysis. For GA, our main concern here, these analyses are based on dioeerent approaches : ffl Proofs of convergence based on Markov chain modeling [6, 3, 1, 20] ffl Deceptive functions analysis, based on Schema analysis and Holland s original theory [14, 8, 9, 11], which characterizes the eOEciency of a GA, and allows to shed light on iGA hardj functions. ffl Some rather new approaches are based on an explicit modelization of a GA as a dynamical system [16, 22] Deception has been intuitively related to the biological notion of epistasis [5] which can be ....

.... nd it. On the contrary, if the intersection of these building blocks is a secondary optimum, the population will preferably converge onto it, missing the global one. In this situation the GA will be considered to have failed 1 and f will be called deceptive. More formally, Goldberg ( 8] [9]) de ned the static deception : The selection results in an expected greater mean tness for the set of individuals selected for reproduction, than for the preceding population. But this mean value will be changed by the application of genetic operators. It follows that the GA can be considered as ....

[Article contains additional citation context not shown here]

David E. Goldberg. Genetic algorithms and walsh functions: Part ii, deception and its analysis. In Complex Systems, volume 3, pages 153171. 1989.

....for a GA and to calculate the average fitness values of schemata. As a basis, these Walsh functions, which take values 1, are more practical than the traditional trigonometric basis like the Fourier functions. A good introduction to using Walsh functions in a GA context is given by Goldberg [18, 19, 20]. Walsh polynomials formalized Walsh functions form an ordered set of rectangular waveforms taking one or two amplitude values: 1 and 1. They form a complete orthogonal set of functions and can thus be used as a basis. Consequently, any fitness function depending on l bits can be thought of as ....

D.E. Goldberg; Genetic Algorithms and Walsh Functions: Part II, Deception and Its Analysis, Complex Systems 3, 153-171, 1989.

....algorithms (MGAs) developed by Goldberg to handle deception, need to identify deceptive schemas to be applicable [13, 14] We suggest an approach satisfying both criteria, using designer genetic algorithms. Deception can be statically identified using the ANODE algorithm suggested by Goldberg [11, 12]. Recent results indicate that the Nonuniform WalshSchema Transform (NWST) 2] can dynamically analyze a GA. Using the NWST in concert with the normal operation of a GA, we can collect runtime statistics needed to identify deception. Furthermore, we can improve efficiency by removing some of the ....

Goldberg, David E., "Genetic Algorithms and Walsh Functions: Part II, Deception and its Analysis, in Complex Systems, 3, 1989, 153-171.

....any complex number a and subsets i and x of a finite universe U we define G a;i (x) a jjinxjj : Here jjsjj denotes the number of elements of the set s, and i nx consists of those elements in i that are not in x. The so called bit products B i (x) and Walsh products R i (x) see for example [7,9,10,15]) are special cases. Given an index set i and a bit string x, a bit product B i (x) multiplies those elements of the bit string x that are in the index set i. Walsh products R i (x) are similar to the bit products, but the bit strings are changed into f Gamma1; 1g strings, and products are taken ....

....on the rate of convergence. In that paper the following restrictions are made: the distributions need to be symmetric, only the fitness function that counts the number of ones in a bit string is considered, and only proportional selection and uniform crossover are treated. In the literature (e.g. [3,9,10]) several applications of Walsh products (but not expected values of Walsh products) can be found in the field of genetic algorithms, for example for the construction of deceptive fitness functions (functions that are difficult for genetic algorithms) for the construction of a number of measures ....

D.E. Goldberg, Genetic algorithms and Walsh functions: Part II, Deception and its analysis, Complex Systems 3 (1989) 153--171.

....Alternatively, there are detailed mathematical formulations of genetic algorithms which do give a detailed description. The two most developed are the Markov Chain formulation of Vose et al. 4] and the Walsh function analysis introduced by Bethke [5] and much used by Goldberg and collaborators [6,7,8]. The former is a beautiful and exact theory of the evolution of a GA. It has been useful in revealing the types of dynamics which can occur (e.g. punctuated equilibria) However, it is of little practical utility, as in order to make predictions, knowledge of transition probabilities between ....

D. E. Goldberg, Genetic Algorithms and Walsh Functions: Part II, Deception and its Analysis, Complex Systems, 3, 153-171, 1989.

....repeat some of the informal arguments used in the previous section, but in a more systematic and rigorous way. Next we consider the case of three classes of problem which have frequently been used in GA research the simple onemax function, the deceptive functions as introduced by Goldberg [8, 9], and the long path problems described by Horn et al. 3] Because the first two classes are already well known, the Root2path problem will be the main example used. Finally, some experimental results will be discussed in the light of the preceding analysis. 2 General Background Traditional ....

....to solve it by a simple steepest ascent hill climber using the Hamming cube operator. However, Culberson [1] shows that the CX landscape for the onemax function is a very different animal. In fact, there is an exponentially increasing (in l) number of local optima. 4. 2 Deceptive problems Goldberg [8, 9] developed the idea of deception from a schema processing perspective. The concept can be simply understood from the 3 bit function in Table 2, which is a version of that originally obtained by Goldberg. This function has the property that while 111 is the global optimum, any schema containing 1s ....

[Article contains additional citation context not shown here]

D.E.Goldberg (1989) Genetic algorithms and Walsh functions: part II, deception and its analysis. Complex Systems, 3, 153-171.

....has come from empirical observations, genetic algorithms are not well understood theoretically. This may be a controversial view, because many theoretical approaches have been proposed. These include static studies of the fitness landscape, such as hyperplane analysis [1] and deception analysis [2], Markov chain analysis [3, 4] and direct methods for small or special problems (e.g. 5] and others. Although these techniques may be useful in providing insight as to how GAs work (or in some cases, it has been argued [6] erroneous insight) it is still difficult to get quantitative results ....

David E. Goldberg. Genetic algorithms and walsh functions: Part II, deception and its analysis. Complex Systems, 3:153--171, 1990.

....characterize problems which GAs are expected to find hard to solve. Such studies have helped to illuminate our understandng of the methodology, while also raising important questions for further research. A particular example of this approach is Goldberg s development of the concept of deception [1, 2]. Problems such as this have been dubbed GA hard although it should not be thought that either all deceptive problems are necessarily hard for a GA to solve, nor that non deceptive problems are easy. Presented at the 15 th Triennial Symposium on Mathematical Programming, Ann Arbor, ....

D.E.Goldberg (1989) Genetic algorithms and Walsh functions: part II, deception and its analysis. Complex Systems, 3, 153-171.

....algorithms (GAs) work still eludes us. One popular approach in the GA community has been to restrict the analysis to a particular class of problems in an endeavour to understand more fundamentally how it is that GAs really work. For example, Goldberg s development of the concept of deception [1, 2], and the Royal Road functions of Mitchell et al. 3, 4] have been useful in helping to understanding the strengths and weaknesses of GAs. In this paper we will consider the case of the well known Onemax problem in some detail. For a binary string x of length l this is the problem of ....

David Goldberg. Genetic algorithms and Walsh functions: Part II, Deception and its analysis. In Complex Systems, volume 3. 1989.

....and GA hardness analysis. For GA, our main concern here, these analyses are based on different approaches : ffl Proofs of convergence based on Markov chain modeling [6] 3] 1] 20] ffl Deceptive functions analysis, based on Schema analysis and Holland s original theory [14] 8] [9], 11] which characterizes the efficiency of a GA, and allows to shed light on GA hard functions. ffl Some rather new approaches are based on an explicit modelization of a GA as a dynamical system [16] 22] Deception has been intuitively related to the biological notion of epistasis [5] ....

....find it. On the contrary, if the intersection of these buildings blocks is a secondary optimum, the population will preferably converge onto it, missing the global one. In this situation the GA will be considered to have failed 1 and f will be called deceptive. More formally, Goldberg ( 8] [9]) defined static deception : the selection results in an expected greater mean fitness for the set of individuals selected for reproduction, than for the preceding population. But this mean value will be changed by the application of genetic operators. It follows that the GA is attracted toward ....

[Article contains additional citation context not shown here]

D.E. Goldberg. Genetic algorithms and walsh functions: Part ii, deception and its analysis. In Complex Systems, volume 3, pages 153--171. 1989.

....Laplace, and other transformations, Walsh functions are often used to represent the representation in a convenient form. Application of Walsh transformation (WT) in understanding Genetic Algorithms was first noted by Bethke [2] Further investigation of this approach can be found elsewhere [4, 5, 6, 13, 16]. Traditionally, the Walsh functions are used for representing real valued functions of binary variables. However, they can be easily extended to higher cardinality representation, as shown elsewhere [15] Although the main arguments of the following discussion can be extended for higher ....

D. E. Goldberg. Genetic algorithms and Walsh functions: Part II, deception and its analysis. Complex Systems, 3(2):153--171, 1989. (Also TCGA Report 89001).

....established that they are inferior. Davis (1991) for example, has argued 1 See, for example, Syswerda (1989) for a description of these 2 nee intrinsic forcefully that on real world problems he invariably gains superior results with natural representations and custom operators, while Goldberg (1990c) has defended the use of binary representations and traditional operators for problems in real parameter optimisation with the development of his theory of virtual alphabets . It should be stressed that there are dozens, if not hundreds of papers which discuss and form part of this debate, and ....

....problem, and is in some ways analogous to the common one max problem (or counting ones problem, e.g. Vose Liepins (1991) 3. 2 Epistatic problems Probably the most widely discussed epistatic problems in the literature on genetic algorithms are the deceptive problems introduced by Goldberg (1990a, 1990b) together with the epistatic members of De Jong s test suite (DeJong, 1975) None of these can be applied directly since they are all defined with respect to arbitrary binary strings, 0 1 2 3 4 5 0 1 2 3 4 5 0 1 2 3 4 5 0 1 2 3 4 5 0 1 2 3 4 5 0 1 2 3 4 5 Figure 2: The left figure shows ....

Goldberg, 1990b. David E. Goldberg. Genetic algorithms and walsh functions: Part II, deception and its analysis.

....has concentrated on three types of features: deception, sampling error, and the ruggedness of a fitness landscape. Bethke [2] defined a class of functions that are misleading for the GA and therefore hard to optimize. Goldberg extended this work, defining the class of GA deceptive functions [7, 8, 10], in which low order schemas lead the GA away from the fittest higher order schemas. There have been a number of studies of GA performance on deceptive landscapes (e.g. 7, 3, 21] Grefenstette and Baker studied a function in which high variance in the fitness of a correct low order schema leads ....

.... hypothesis [14] is that the GA should be better able to search landscapes containing such features because the lower fitness deserts can be quickly crossed via crossover (here, between instances of 11 and 11) Isolates are a special case of what have been called partially deceptive functions [8]. The idea of isolated regions of high fitness surrounded by flat deserts of low fitness is similar to the mesa phenomenon proposed by Minsky [24] and to the error surfaces identified by Hush et al. for multilayer perceptron neural networks [16] Thus, the shape of the surface may be as ....

D. E. Goldberg. Genetic algorithms and Walsh functions: Part II, Deception and its analysis. Complex Systems, 3:153--171, 1989.

No context found.

Goldberg, D. E., 1989a, Genetic algorithms and Walsh functions: Part II, deception and its analysis, Complex Syst. 3:153--171.

.... are sufficient to unambiguously identify a unique D, which is the complement of g (Whitley, 1991) We can then construct fully deceptive functions in which the schema containing D is the winner of every partition (i.e. types II and III deception) at every order up to the string length (Goldberg, 1989a, 1989b, 1990). Full deception is clearly maximally misleading to a GA. It is also clearly bimodal, with local optima 21 at D and at g. To add more optima, and basins of GA attraction, we need to define partial deception. Deception becomes a more practical tool of GA theory when it is embedded in fitness ....

Goldberg, D. E. (1989c). Genetic algorithms and Walsh functions: part II, deception and its analysis. Complex Systems, 3, 153--171.

....has focused on deception. For example, there have been works that have tried to better define or classify deception (Goldberg, 1987, 1989a, 1989b; Liepins Vose, 1990; Vose Liepins, 1991; Whitley, 1991) studies that have constructed examples of deceptive functions (Deb Goldberg, 1991, 1992; Goldberg, 1990; Goldberg, 1989a; Whitley, 1991) and investigations that have altered genetic algorithms to try to solve deceptive problems (Goldberg, Deb, Korb, 1990; Goldberg, Korb, Deb, 1989; Eshelman, 1991) Despite these and many other advances, questions have been raised about what deception is about ....

....a set of sufficient conditions for deception in such functions. In the next section, we will review the construction of two examples of such functions, one modeled after the bimodal trap functions and the other modeled after deceptive functions constructed from order limited Walsh coefficients (Goldberg, 1990). 2.1 Defining bipolar deception In the usual deceptive functions, a schema partition is defined to be deceptive if the schema containing the deceptive optimum is better than any other competing schema in the partition. In a bipolar function, there exist two global optima and a number of ....

[Article contains additional citation context not shown here]

Goldberg, D. E. (1989b). Genetic algorithms and Walsh functions: Part II, deception and its analysis.

No context found.

Goldberg, D. E.: Genetic Algorithms and Walsh Functions: Part II, Deception and Its Analysis. Complex Systems (1989) 153-171

No context found.

Goldberg, D. E.: Genetic algorithms and Walsh functions: part II, deception and its analysis, Complex Systems (1989) 153-171

No context found.

D. E. Goldberg. Genetic algorithms and Walsh functions: Part II. Deception and its analysis. Complex Systems, 3:153--171, 1989.

No context found.

Goldberg, D. E. (1989b). Genetic algorithms and Walsh functions: Part II, deception and its analysis.

No context found.

Goldberg, D. E. (1989c). Genetic algorithms and Walsh functions: Part II, deception and its analysis. Complex Systems, 3, 129-152.

First 50 documents  Next 50

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC