|
NEU-2003
Optimal Control with Genetic Algorithms
Andrei V. Kvichansky, Serge A. Terekhoff
Artificial Neural Networks Laboratory,
NeurOK Software, LLC, Moscow
Important class of controlled systems – modern bioreactors
– are operating in fed-batch mode. During the batch run, several phases
with different substrate consumption and metabolic production rates are
distinguished. As a result, the states of variables may vary significantly
during the operation. Optimal policy of feeding schedule of the reactor is
a great scientific and technological challenge.
Feeding profile can be sought as the solution of
multidimensional optimization problem, where control values on each time
step are treated as unknowns.
Genetic Algorithms (GA) offers robust global optimal
solution for multivariate problems. The solution (for example, a schedule)
is encoded in a chromosome representation. GA uses Darwinian principles of
random mutation and crossover to improve solution to a problem. The
solutions in a population combine (crossover) and change randomly (mutate)
during each generation of offspring.
The principal scheme of GA solution to a control problem
is presented below.
In this report we investigate two models of industrial
plants comprising systems of ordinary differential equations, and build GA
solutions of variation problems for them.
The method is implemented in C++ software module, which
uses the GAlib library.
|
