I used a GA to optimize seating assignments at my wedding reception. 80 guests over 10 tables. Evaluation function was based on keeping people with their dates, putting people with something in common together, and keeping people with extreme opposite views at separate tables.
I ran it several times. Each time, I got nine good tables, and one with all the odd balls. In the end, my wife did the seating assignments.
My traveling salesman optimizer used a novel mapping of chromosome to itinerary, which made it trivial to breed and mutate the chromosomes without any risk of generating invalid tours.
Update: Because a couple people have asked how ...
Start with an array of guests (or cities) in some arbitrary but consistent ordering, e.g., alphabetized. Call this the reference solution. Think of a guest's index as his/her seat number.
Instead of trying to encode this ordering directly in the chromosome, we encode instructions for transforming the reference solution into a new solution. Specifically, we treat the chromosomes as lists of indexes in the array to swap. To get decode a chromosome, we start with the reference solution and apply all the swaps indicated by the chromosome. Swapping two entries in the array always results in a valid solution: every guest (or city) still appears exactly once.
Thus chromosomes can be randomly generated, mutated, and crossed with others and will always produce a valid solution.
