Global+Optimization+Algorithms+Theory+and+Application_Part16

Global+Optimization+Algorithms+Theory+and+Application_Part16...

Info iconThis preview shows pages 1–3. Sign up to view the full content.

View Full Document Right Arrow Icon

Info iconThis preview has intentionally blurred sections. Sign up to view the full version.

View Full DocumentRight Arrow Icon
This is the end of the preview. Sign up to access the rest of the document.

Unformatted text preview: 18.2 Distribution 301 local machine Initial. Eval. Fitness. Select. Reprod. Initial. Eval. Reprod. main thread worker thread worker thread Figure 18.3: A parallel evolutionary algorithm with two worker threads. Cant’u-Paz [328] divides parallel evolutionary algorithms into two main classes: 1. In globally parallelized EAs, each individual in the population can (possibly) always mate with any other. 2. In coarse grained approaches, the population is divided into several sub-populations where mating inside a sub-population is unrestricted but mating between individuals of different sub-populations may only take place occasionally according to some rule. In ancient Greece, a deme was a district or township inhabited by a group that formed an independent community. They were the basic units of government in Attica as remodeled by Cleisthenes around 500 BC. In biology, a deme is a locally interbreeding group within a geographic population. Definition 18.1 (Deme). In evolutionary algorithms, a deme is a distinct sub-population. In the following, we are going to discuss some of the different parallelization methods from the viewpoint of distribution because of its greater generality. 18.2 Distribution The distribution of an algorithm only pays off if the delay induced by the transmissions necessary for data exchange is much smaller than the time saved by distributing the com- putational load. Thus, in some cases distributing of optimization is useless. If searching for the root of a mathematical function for example, transmitting the parameter vector x to another computer will take much longer than computing the function f ( x ) locally. In this section, we will investigate some basic means to distribute evolutionary algorithms that can as well as be applied to other optimization methods as outlined by Weise and Geihs [2177]. 18.2.1 Client-Server If the evaluation of the objective functions is time consuming, the easiest approach to dis- tribute and evolutionary algorithm is the client-server scheme (also called master-slave). 9 Figure 18.4 illustrates how we can make use of this very basic, global distribution scheme. Here, the servers (slaves) receive the single tasks, process them, and return the results. 9 A general discussing concerning the client-server architecture can be found in Section 30.2.2 on page 556 302 18 Parallelization and Distribution Such a task can, for example, be the reproduction of one or two individuals and the sub- sequent determination of the objective values of the offspring. The client (or master) just needs to distribute the parent individuals to the servers and receives their fully evaluated offspring in return. These offspring are then integrated into the population, where fitness assignment and selection is performed. Client-server-based distribution approaches for evo- lutionary algorithms have been discussed and tested by Van Veldhuizen et al. [2102], Xu et al. [2271], Dubreuil et al. [604] and were realized in general-purpose software packages byet al....
View Full Document

Page1 / 20

Global+Optimization+Algorithms+Theory+and+Application_Part16...

This preview shows document pages 1 - 3. Sign up to view the full document.

View Full Document Right Arrow Icon
Ask a homework question - tutors are online