Heinz M uhlenbein, Thilo Mahnig, and Aberto O. Rodriguez. Schemata, distributions and graphical models in evolutionary optimization. J. of Heuristics, 5:215--247, 1999.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....Evolutionary Computation (EC) are widely used to tackle hard optimization problems. More recently we can highlight new population based models that rely on probability distributions that sample the search space [1] or that take into account the variable dependencies like Bayesian networks [6, 7]. In this paper we show that another kind of bio inspired algorithm, novel ant based algorithm, can be used to perform probabilistic search with the help of Hidden Markov Models (HMMs) Ant based methods share a common feature with EC: a population of solutions is maintained. The population used ....

H. Miihlenbein, T. Mahnig, and A. Ochoa. Schemata, distribution and graphical models in evolutionary computation. Journal of Heuristics, 5(2):215-247, 1999.

....optimal solution grows exponentially with the problem size in absence of the knowledge regarding the important partitions. The genetic algorithm community have realized the importance of linkage learning in the scalability of the genetic algorithm. A growing number of linkage learning techniques [3, 4, 12, 23, 20, 24, 32, 31, 41, 54, 55, 56, 70] are becoming available. The following part of this section presents a brief review of some of the earlier e orts in linkage learning. The history of linkage learning e orts dates back to Bagley s dissertation [2] Bagley used a exible representation and the so called inversion operator for ....

....approach also exploits the second order non linearity and estimates a second order approximation of the underlying relations; they reported superior performance of their algorithm over other techniques that do not explicitly try to search and exploit relations. The approach proposed elsewhere [54, 55] is based on estimation of the underlying distribution function. However, since distribution estimation is a hard problem faced in many domains, their approach assumes prior knowledge about the structure of the distribution and reduces the task to estimation of the distribution parameters with ....

H. Muhlenbein and A. O. Rodriguez. Schemata, distributions and graphical models in evolutionary optimization. Personal Communication., December 1997.

....and M uhlenbein [6] and De Bonet et al. 7] A step forward involved to factorize the joint probability distribution, in a tree like structure, 8] 9] and [10] Recently works have appeared where the joint probability distribution is factorized by means of Bayesian networks. M uhlenbein et al. [11] have developed an algorithm for additive decomposable functions called FDA. The more general approaches have been developed by Etxeberria and Larra naga [12] and Pelikan et al. 13] In spite of this e ort to obtain new algorithms, few attention have received the theoretical aspects of them. One ....

Muhlenbein, H., Mahning, T. & Ochoa, A. Schemata, Distributions and Graphical Models in Evolutionary Optimization. Journal of Heuristics, 5, 215-247, 1999.

....to section 3. More recently, approaches allowing multivariate interactions were presented. In these approaches, a variable may be conditionally dependent on sets of variables. These methods include the BOA by Pelikan, Goldberg and Cant u Paz [24] the FDA by M uhlenbein, Mahnig and Rodriguez [22] and the ECGA by Harik [15] Even though nding the correct structure to capture the sets of interacting variables is very dicult, covering multivariate interactions is the key to solving higher order building block problems and exploiting problem structure [7] M uhlenbein, Mahnig and Rodriguez ....

.... and the ECGA by Harik [15] Even though nding the correct structure to capture the sets of interacting variables is very dicult, covering multivariate interactions is the key to solving higher order building block problems and exploiting problem structure [7] M uhlenbein, Mahnig and Rodriguez [22] rst presented a general framework for this type of algorithm, named EDA (Estimation of Distribution Algorithm) This framework is a general optimization algorithm that estimates a distribution every iteration, based upon a collection of solutions and subsequently samples new solutions from the ....

[Article contains additional citation context not shown here]

H. Muhlenbein, T. Mahnig, and O. Rodriguez. Schemata, distributions and graphical models in evolutionary optimization. Journal of Heuristics, 5:215-247, 1999.

....task of nding the best structure that suits the model is higher. Therefore, these approaches require a more complex learning process. The most important EDA approaches that can be found in the literature within this category are as follows: FDA (Factorized Distribution Algorithm) introduced in [32], EBNA (Estimation of Bayesian Networks Algorithm) 33] BOA (Bayesian Optimization Algorithm) 34] LFDA (Learning Factorized Distribution Algorithm) introduced in [35] that follows essentially the same approach as in EBNA, and the Extend compact Genetic Algorithm (EcGA) proposed in [36] 4 ....

H. Muhlenbein, T. Mahning, and A. Ochoa. Schemata, distributions and graphical models in evolutionary optimization. Journal of Heuristics, 5:215{ 247, 1999.

.... Recently, a number of methods have been proposed that explicitly model the population of good solutions and use the constructed model to guide further search [2, 5, 9, 11] These methods are generally known as the estimation 2 Byoung Tak Zhang and Soo Yong Shin of distribution algorithms or EDAs [8]. They use global information contained in the population, instead of using local information through crossover or mutation of individuals. From the population, statistics of the hidden structure are derived and used when generating new individuals. One of the main issues in distribution based ....

....One of the main issues in distribution based optimization is how to build and sample from the distribution of the population. Several methods have been proposed, including the methods based on dependency chains [5] dependency trees [2] factorization [9] neural trees [13] Bayesian networks [8, 12], and genetic programs [14] In this paper, we present a method that estimates the sample distribution using a graphical learning model known as Helmholtz machines. The method is implemented as a Bayesian evolutionary algorithm (BEA) a probabilistic model of evolutionary computation [13] The ....

[Article contains additional citation context not shown here]

Muhlenbein, H., Mahnig, T., and A. Ochoa, \Schemata, distributions and graphical models in evolutionary optimization", Journal of Heuristics, 5:215-247, 1999.

....Covering pairwise interactions still does not preserve partial solutions of higher order. Moreover, interactions of higher order do not necessarily imply pairwise interactions that can be detected at the level of partial solutions of order two. The factorized distribution algorithm (FDA) M uhlenbein et al. 1998) uses a xed factorization of the distribution to generate new candidate solutions. The FDA is capable of covering the interactions of higher order and combining important partial solutions e ectively. It works very well on uniformlyscaled additively decomposable problems. However, the FDA ....

Muhlenbein, H., Mahnig, T., & Rodriguez, A. O. (1998). Schemata, distributions and graphical models in evolutionary optimization. Submitted for publication.

No context found.

Muhlenbein, H., Mahnig, Th., & Ochoa, A. Rodriguez. 1999. Schemata, Distributions and Graphical Models in Evolutionary Optimization. Journal of Heuristics, 5, 215-247.

No context found.

H. Muhlenbein, Th. Mahnig, and A. Rodriguez Ochoa. Schemata, distributions and graphical models in evolutionary optimization. Journal of Heuristics, 5:215-247, 1999.

No context found.

H. Muhlenbein, Th. Mahnig, and A. Rodriguez Ochoa. Schemata, distributions and graphical models in evolutionary optimization. Journal of Heuristics, 5(2):215-247, 1999.

No context found.

M uhlenbein, H., Mahnig, T., and Ochoa, A. (1999). Schemata, distributions and graphical models in evolutionary optimization. Journal of Heuristics, 5(2):213--247.

No context found.

H. Muhlenbein, Th. Mahnig, and A. Rodriguez Ochoa. Schemata, distributions and graphical models in evolutionary optimization. Journal of Heuristics, 5:215{ 247, 1999.

No context found.

Muhlenbein, H., T. Mahnig, and A. R. Ochoa (1999). Schemata, distributions and graphical models in evolutionary optimization. Journal of Heuristics 5, 215-247.

No context found.

H. M uhlenbein, T. Mahnig, and A. Ochoa. Schemata, distributions and graphical models in evolutionary optimization. Journal of Heuristics, 5(2):213--247, 1999.

.... f(x) p(y)e f(y) 17) We can now de ne the BEDA (Boltzmann Estimated Distribution Algorithm) BEDA is a conceptual algorithm, because the calculation of the distribution requires a sum over exponentially many terms. We have proven the following important convergence theorem for it [24]. Theorem 13 (Convergence) Let (t) be an annealing schedule, i.e. for every t increase the inverse temperature by (t) Then for BEDA the distribution at time t is given by (t)f(x) Z f ( t) 18) with the inverse temperature (t) 1 ( 19) Algorithm 2: BEDA Boltzmann ....

....i = 1; m the sets d i , b i and c i : d i : s j ; b i : s i n d i 1 ; c i : s i d i 1 (22) We set d 0 = In the theory of decomposable graphs, d i are called histories, b i residuals and c i separators [16] We now need the conditional probabilities from de nition 2. In [24] we have proven the following theorem. Theorem 16 (Factorization Theorem) Let p (x) be a Boltzmann distribution with (23) and f(x) P m f s i (x) be an additive decomposition. If b i 6= 8i = 1; l; d l = X; 24) 8i 2 9j i such that c i s j (25) then Ym p(x b i jx c i ....

H. Muhlenbein, Th. Mahnig, and A. Rodriguez Ochoa. Schemata, distributions and graphical models in evolutionary optimization. Journal of Heuristics, 5(2):215-247, 1999.

....y i g. In order to derive Wright s equation, we have to introduce a special distribution. De nition: Robbins proportions are de ned by the distribution p (x; t) p i (x i ; t) 2) A population in Robbins proportions is called to be in linkage equilibrium in population genetics. In [7, 11] we have shown: All complete recombination schemes lead to the same univariate marginal distributions after one step of selection and recombination. If recombination is used for a number of times without selection, then the genotype frequencies converge to linkage equilibrium. This means that all ....

H. Muhlenbein, Th. Mahnig, and A. Rodriguez Ochoa. Schemata, distributions and graphical models in evolutionary optimization. Journal of Heuristics, 5:215-247, 1999.

No context found.

) Muhlenbein, H. & Mahnig, Th. & Ochoa, R. (1999a). Schemata, Distributions and Graphical Models in Evolutionary Optimization. Journal of Heuristics,

....In order to derive Wright s equation, we have to introduce a special distribution. De nition: Robbins proportions are de ned by the distribution p (x; t) n Y i=1 p i (x i ; t) 2) A population in Robbins proportions is called to be in linkage equilibrium in population genetics. In [3, 6] we have shown: All complete recombination schemes lead to the same univariate marginal distributions after one step of selection and recombination. If recombination is used for a number of times without selection, then the genotype frequencies converge to linkage equilibrium. This means that all ....

....of UMDA. The mean tness increases with z. The extension of the above lemmata to multiple alleles and multivariate distributions is straightforward, but the notation becomes dif cult. Multivariate distributions are used by an extension of UMDA, the Factorized Distribution Algorithm FDA [5, 6] 4 The Selection Problem Fitness proportionate selection is the undisputed selection method in population genetics. It is considered to be a model for natural selection. But this selection method strongly depends on the tness values. When the population approaches an optimum, selection gets ....

H. Muhlenbein, Th. Mahnig, and A. Rodriguez Ochoa. Schemata, distributions and graphical models in evolutionary optimization. Journal of Heuristics, 5:215{ 247, 1999.

....describe the general context. In our research we have transformed evolutionary algorithms which recombine strings to algorithms which use search distributions. For a conceptual algorithm (Boltzmann distribution estimation algorithm BEDA) we have proven convergence to the set of global optima [MMO99]) BEDA STEP 0: Set t ( 1. Generate N 0 points randomly. STEP 1: Select M N points according to Boltzmann selection. Estimate the distribution p s (x; t) of the selected set. STEP 2: Generate N new points according to the distribution p(x; t 1) p s (x; t) Set t ( t 1. ....

....is given by T (x) 1 Z T exp f(x) T (1) Z T is the usual partition function, de ned by P x exp f(x) T . It can easily be shown that the search distribution of the selected points p s (x; t) is a Boltzmann distribution if the total distribution p(x; t) is a Boltzmann distribution [MMO99]. From BEDA a practical algorithm has been derived, called the Factorized Distribution Algorithm (FDA) It assumes that the search distribution can be factored into a Bayesian network p s (x; t) Q i p(x i jpa i ) pa i are called the parents of x i . The interested reader is referred to ....

[Article contains additional citation context not shown here]

H. Muhlenbein, Th. Mahnig, and A. Rodriguez Ochoa. Schemata, distributions and graphical models in evolutionary optimization. Journal of Heuristics, 5:215-247, 1999.

No context found.

) Muhlenbein, H. & Mahnig, Th. & Ochoa, R. (1999a). Schemata, Distributions and Graphical Models in Evolutionary Optimization. Journal of Heuristics,

....where # 0 is a parameter, also called the inverse temperature, and W # = # p(x, t )e #f (x) is the weighted average of the population. For Boltzmann distributions we have proven a factorization theorem for the distribution p(x, t ) and convergence for an algorithm using this factorization [20]. The proof is simple, because if p(x, t ) is a Boltzmann distribution with factor # 1 and Boltzmann selection is done with factor # 2 , then p(x, t 1) p(x, t ) is a Boltzmann distribution with factor # = # 1 # 2 . Theorem 9. Let p(x, 0) be randomly distributed. Let # 1 , # ....

....should be the first algorithm to be tried in a practical algorithm. In a next step all three algorithms have to be extended to optimization problems with constraints. We believe that with distributions constraints can be easier handles than with recombination. A first step has already been made in [20]. ....

H. Mhlenbein, Th. Mahnig, and A. Rodriguez Ochoa. Schemata, distributions and graphical models in evolutionary optimization. Journal of Heuristics, 5:215--247, 1999.

No context found.

Heinz M uhlenbein, Thilo Mahnig, and Aberto O. Rodriguez. Schemata, distributions and graphical models in evolutionary optimization. J. of Heuristics, 5:215--247, 1999.

No context found.

M uhlenbein, H., Mahnig, T., and Rodriguez, A. O. (1999). Schemata, distributions and graphical models in evolutionary optimization. J. of Heuristics, 5:215--247.

No context found.

H. M uhlenbein, T. Mahnig, and A. O. Rodriguez. Schemata, distributions and graphical models in evolutionary optimization. Journal of Heuristics, 5:215--247, 1999.

No context found.

H. Muhlenbein, T. Mahnig, and A. Ochoa (1999). Schemata, Distributions and Graphical Models in Evolutionary Computation. Journal of Heuristics 5, 215-247.

First 50 documents

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