Genetic algorithms are multiparametric procedures. Their operations depend on a variety of parameters these usually are the size of the initial population, the number of generations, probability thresholds that define the crossover and mutation procedures and sometimes heuristics that some developers use. One of the open problems on genetic algorithms is the determination of the optimal values for these parameters. It is also called parameter tuning.
Like genetic algorithms, there are other multiparametric were the determination of the optimal values for their parameters is crucial for their operation. In all the above cases, sensitivity analysis is used for reaching optimality.
As parameter tuning is actually an optimization problem and genetic algorithms are optimization techniques, I thought it would be interesting to build a genetic algorithm for parameter tuning. The result of this work may be found here and it is released as an open source project.
The tuning algorithm examines the program that implements some multiparametric algorithm as a black box. The internal operations of the program are not examined; it only considers the output of the program given some valuation on its parameters. The population of the tuning algorithm consists of set of such valuations. It is an effective approach as I have tested it extensively.
It is a very interesting approach as we do not have to consider the functionality of the program under study, as the tuning algorithm adapts its functionality on the program. And this is the essence of artificial intelligence; the algorithms have to adapt to their subject of study.
When the program under study implements a genetic algorithm then we have a genetic algorithm that tunes up another genetic algorithm, which is a cool idea.
Of course the tuning algorithm is also multiparametric, but in it we do not seek optimality. It is enough to set some high values on its parameters, wait some time until it terminates and compute the optimal values for the program. Then run the program with optimal performance.
Like genetic algorithms, there are other multiparametric were the determination of the optimal values for their parameters is crucial for their operation. In all the above cases, sensitivity analysis is used for reaching optimality.
As parameter tuning is actually an optimization problem and genetic algorithms are optimization techniques, I thought it would be interesting to build a genetic algorithm for parameter tuning. The result of this work may be found here and it is released as an open source project.
The tuning algorithm examines the program that implements some multiparametric algorithm as a black box. The internal operations of the program are not examined; it only considers the output of the program given some valuation on its parameters. The population of the tuning algorithm consists of set of such valuations. It is an effective approach as I have tested it extensively.
It is a very interesting approach as we do not have to consider the functionality of the program under study, as the tuning algorithm adapts its functionality on the program. And this is the essence of artificial intelligence; the algorithms have to adapt to their subject of study.
When the program under study implements a genetic algorithm then we have a genetic algorithm that tunes up another genetic algorithm, which is a cool idea.
Of course the tuning algorithm is also multiparametric, but in it we do not seek optimality. It is enough to set some high values on its parameters, wait some time until it terminates and compute the optimal values for the program. Then run the program with optimal performance.
