I could give some implementation detauils, I assume...
The NN is simply a map\, float> , representing the connections between nodes in the network. (Template parameters beeing From, To and Weight.)
Currently, I started from zero and randomly add connections and nodes, modify weights etc. in the mutation phase.
The reason I wrote this is because it seemed like fun.
I don't see how I could train the network manually, since I have no idea what the topology of the network should look like. I believe I need at least a few thousand connections and hundreds of middle layer nodes in the network. Shouldn't this take looong time to train manually?
Currently my NNs play against eachother, randomly. But since the same rules always apply, shouldn't they learn from that?
I have run the evolution for approx. 1000 generations now (on my slow 266mhz laptop) with a population of 50 NNs. They seem to be playing the opening game pretty good now, but they are still crappy in the end of the matches. (I play against them regularly to try them out.)