A
GA that employs
Lamarckian inheritance, rather than
Darwinian. That is, each
string (coding for an individual in the population) passes on some of the characteristics it has acquired during its lifetime. Since many GAs operate a breed 'n die policy, this is only relevant for programs that do something other than just evaluate a fitness function.
Since this is what I know best, I will describe a molecular LGA. The xyz coordinates of each conformer are encoded as a string (say 10.20.30 for a really simple coding of one atom). Random mutations are made to the strings and crossovers between them during breeding. During the 'life' of a conformer, it may do some local movements that can be transmitted to its offspring. Since the local search alters the strings phenotype, and is then recorded into its genotype this is lamarckian.