Parallel Evolutionary Optimization under Matlab on standard computing networks

For many computing intensive tasks at DaimlerChrysler Research MATLAB [1] is used. Matlab has become a defacto standard in research and development. Many areas are catered for by a wide range of toolboxes and the Simulink non-linear simulation package along with extensive visualization and analysis tools. In addition, Matlab has an open and extensible architecture allowing individual users to develop further routines for their own applications. These qualities provide a uniform and familiar environment on which to use and build extended tools. Since 1995 we use Evolutionary Algorithms for many difficult optimization tasks. We employ the Genetic and Evolutionary Algorithm Toolbox for use with Matlab (GEATbx) [6]. This toolbox provides a large number of different operators and functions for a wide range of evolutionary algorithms. For most evolutionary optimization tasks the execution times were satisfying on standard PC's or workstations (some minutes to 1-2 hours). However, when optimizing large systems one run could take more than a day. Such a long optimization time is not acceptable. From experiments with test systems we knew, that the evolutionary part of an optimization run takes only 2-10 minutes. The remaining time is spent on executing the objective function. Evolutionary Algorithms are inherent parallel. At each generation the same objective function is executed with multiple different data sets (the individuals). This "single program multiple data" scheme opens the way for shorter optimization times by distributing the calculation of the objective function to multiple machines. Multi processor machines and clusters of networked computers are readily available today, especially networks of standard PC's and UNIX/LINUX workstations. In 1997 we looked for a tool providing the functionality of distributing computing tasks to multiple machines or processors. Our aim was to include this functionality into Matlab in a transparent way. We appreciate using our familiar and powerful tools, but combined with a comfortable way of distributing calculations to multiple machines.