Differences
This shows you the differences between two versions of the page.
|
rudolph:ea [2010-04-08 19:17] |
rudolph:ea [2010-06-30 16:37] |
||
|---|---|---|---|
| Line 4: | Line 4: | ||
| For approximately 4 billion years, there has been life on earth. Apart from the amazing diversity of species, the process of evolution has created many organisms and forms that are well adapted to their respective environment, partly even in an optimal way. Why should one not try to come to new and more robust optimization procedures by mimicking fundamental evolutionary principles? | For approximately 4 billion years, there has been life on earth. Apart from the amazing diversity of species, the process of evolution has created many organisms and forms that are well adapted to their respective environment, partly even in an optimal way. Why should one not try to come to new and more robust optimization procedures by mimicking fundamental evolutionary principles? | ||
| + | {{:rudolph:ea_ablauf.png?500 |Workflow of EAs}} | ||
| At the beginning of the 1960s, different researchers came up with this question independently of each other. In Germany, this has led to evolution strategies (ES), in the USA to genetic algorithms (GA) and the concept of evolutionary programming (EP). These procedures as well as genetic programming (GP), which transfers evolutionary principles into the search space of programming languages, are summarized today under the names evolutionary algorithms (EA) or evolutionary computation (EC). The different classes of evolutionary algorithms differ by the representation of the individuals and by their variation and selection operators. An important advantage of all evolutionary procedures is their inherent, scalable parallelism. Thus, EA can easily be adapted to any kind of parallel data processing architecture. | At the beginning of the 1960s, different researchers came up with this question independently of each other. In Germany, this has led to evolution strategies (ES), in the USA to genetic algorithms (GA) and the concept of evolutionary programming (EP). These procedures as well as genetic programming (GP), which transfers evolutionary principles into the search space of programming languages, are summarized today under the names evolutionary algorithms (EA) or evolutionary computation (EC). The different classes of evolutionary algorithms differ by the representation of the individuals and by their variation and selection operators. An important advantage of all evolutionary procedures is their inherent, scalable parallelism. Thus, EA can easily be adapted to any kind of parallel data processing architecture. | ||
| Line 11: | Line 12: | ||
| ===== Evolution strategies ===== | ===== Evolution strategies ===== | ||
| - | Although invented originally for experimental optimization and discrete search spaces, typically the is the search domain of evolution strategies since their implementation on computers. For this case, the theory of ES progressed furthest. Accordingly, an individual consists of an n- dimensional real valued vector //x// representing the object variables and a number of strategy parameters controlling mutation. The number of strategy parameters can vary between 1 and //n + n(n 1)/2// depending on the user's choice. | + | Although invented originally for experimental optimization and discrete search spaces, typically the real numbers are the search domain of evolution strategies since their implementation on computers. For this case, the theory of ES progressed furthest. Accordingly, an individual consists of an //n//-dimensional real valued vector //x// representing the object variables and a number of strategy parameters controlling mutation. The number of strategy parameters can vary between 1 and //n// + //n//(//n// - 1)/2 depending on the user's choice. |
| The special feature of ES is recombining and mutating both object and strategy parameters. This and an appropriate selection pressure provide the possibility of a permanent self-adaptation of mean step sizes, and sometimes even the preferred directions the strategy takes in the search space emerge. | The special feature of ES is recombining and mutating both object and strategy parameters. This and an appropriate selection pressure provide the possibility of a permanent self-adaptation of mean step sizes, and sometimes even the preferred directions the strategy takes in the search space emerge. | ||