An optimization problem as complex as residual statics estimation in s
eismic image processing requires novel techniques. One interesting tec
hnique, the genetic algorithm, is based loosely on the optimization pr
ocess forming the basis of biological evolution. The objective of this
paper is to examine this algorithm's applicability to residual static
s estimation and present three new ingredients that help the algorithm
successfully resolve residual statics. These three ingredients includ
e (1) breaking the population into subpopulations with restricted bree
ding between the subpopulations, (2) localizing the search, to varying
degrees, about the uncorrected input stack, and (3) modifying the opt
imization function to take account of CDP-dependent structural feature
s. Introducing subpopulations has the effect of enhancing the search w
hen the volume of phase space being searched is large and limited info
rmation is given about where the algorithm should concentrate its effo
rts. Subpopulations work well initially, but as the performance increa
ses, the search efficiency decreases. Thus, search efficiency is depen
dent on both the subpopulation size and the present performance of the
subpopulation. The greediness of genetic algorithms in their rapid ac
ceptance of a local minimum can be recompensed through a more cautious
and user-controlled exploration of the phase space. Specifically, gen
etic algorithms can be ''fed'' the uncorrected input stack as a way of
biasing the search in the region of phase space that contains the rou
gh event images observable in most uncorrected seismic stacks. We comp
are three types of starting populations: (1) a randomized population,
(2) a biased start with a randomized population save one individual re
presenting the input stack, and (3) a biased start restricted to a slo
wly expanding (generation-dependent) volume of phase space. Efficient
searches also require an optimization function that places the perfect
ly corrected seismic image at the function's global maximum. Construct
ing such a function is nontrivial, and we implement a seismic data set
that allows us to examine the genetic algorithm's sensitivity to inap
propriate optimization functions.