All the elements you mention combined restricts what your project has to be.

Genetic Programming would be something you do at the coarse level because a Neutral Network becomes irrelevant when the problem it is solving changes too much . I could then see the GA setting up the parameters/coefficients defining a Neural Network 'build' (the node layer definition, the training coefficients, the situational interpretation rules/the good/bad result judgement functions, training set selection), which then is applied to build a candidate NN training session (building a NN doing the usual reinforced learning method ). A 'swarm' (swarm learning) of candidates are done with 'genetic' variations and a subset of the 'best' resulting NN then selected as the basis of the next iteration set of GA variations (working toward some 'bestest' resulting NN)

How much you can automate building the training sets depends on the game mechanics of the simulation. (whether it can be subjected to the GA methodology) Training sets can be quite large to cover the spectrum of cases the NN is expected to handle. If they have to be hand built that is a seperate problem in the project (maybe GA can be used to assist in the training set building as an independant part of this project)

Swarm intelligence GA variation control factors itself is something that might be manipulated at a super coarse level.

Get ready for some huge amount of processing - building just one NN can be a mass of computing, and multiple iterations of GA variations multiplies that by magnitudes.

How about doing one of those things at a time instead of trying to combine them all into one master project?