### ASA 125th Meeting Ottawa 1993 May

## 5aAO9. Global inversion using genetic algorithms.

**P. Gerstoft
**

**
**
*Saclant Undersea Res. Ctr., La Spezia, Italy
*

*
*
An inversion of sound fields for determining unknown environmental
parameters can be separated into four parts: (1) discretization of the
environment and discretization or transformation of the data, (2) efficient and
accurate forward modeling, (3) efficient optimization procedures, and (4)
uncertainty analysis. While much work had been done on the first two, much less
has been done on the latter two, especially for object functions with several
local minima. Global optimization methods accept the multiple minima and try to
find the global minimum, without doing an exhaustive search. These methods are
based on directed Monte Carlo search, and two promising methods are simulated
annealing and genetic algorithms (GA). These have been compared and it has been
found that GA's often are superior. The global methods are time demanding and
in order to speed up the convergence, gradient steps can be taken during the
iteration process. The examples here will be based on a horizontally stratified
environment where all the material and geometrical parameters can be taken as
unknown parameters. The solution is presented in terms of a posteriori
probability function describing the parameters. From this the most likely model
parameters can be found and their uncertainty and importance can be assessed.