The genetic algorithm (GA) is extended to solve a family of optimization problems differing in the value of some parameters. The goal of the optimization is thus not to find the optimal solution of a single problem, rather to find the optimal mapping from the parameter space (input space) to the space of the solutions (output space), in other words to find the optimal input-output mapping. The input values are assumed to be continuously varying. The input space is discretized in an optimal fashion with the help of a topology conserving neural network. The GA is generalized to organize individuals into subpopulations associated with the discretization points. Using the topology defined by the discretization method the individuals are allowed to migrate to neighboring sites, while within each subpopulation the original GA operators are applied as the means of evolution. The migration method speeds up the optimization by allowing small subpopulations without losing the diversity within the subpopulations. To illustrate the method the optimal control of a simulated robot-arm is treated: ping-pong balls dropped from varying heights have to be caught by a bat without bouncing, with the dropping height serving as the input. It is demonstrated that the simultaneous optimization for an interval of dropping heights can be solved by GA with migration. Several effects of migration are examined in detail.
|Number of pages||11|
|Journal||Neural Network World|
|Publication status||Published - Jan 1 1995|
ASJC Scopus subject areas
- Hardware and Architecture
- Artificial Intelligence