The Cellz competition received a total of twenty entries in the on-line league, though several of these were test entries by the organiser.
The winner was Jason Brownlee of Victoria, Australia, with his entry JB_Smart_Function_v1.1. Jason tells me he designed the overal neural architecture by hand, taking inspiration from the hand-coded Java controller that was supplied in the developer kit, then used ECJ to tune the weights; the result is a neural controller with very similar behaviour to the hand-designed controller. Can anyone beat this? The on-line league will continue to run, so we'll wait and see!
The state of the on-line league at the closing date of the competition is here.
You can run Jason's entry by downloading the text-file that describes the neural net, which is winner.txt - you also need the Cellz jar file, and a recent version of Java (e.g. Java 1.4).
Then type:
java -classpath cellz.jar games.cellz.DishController winner.txt
Note that this is a reasonably fit solution for the 1,000 time-steps that the evaluation was done for, but suffers the same stagnation problems as the hand-designed controller after a few thousand time-steps. A screen-shot is shown below: