The easiest way to add complexity to the Fitness Landscape is to build some - say 10 to 20 - analog computers, cross-linkable with A/D and D/A conversion to/from the neural net to simulate continuous data streams which the neural net nodes have to 'chunk' - both for data processing and across time.

Analog computers with >90dB dynamic range working with 16 bit A/Ds and D/As could give really fine levels of resolution and speed.  And your point about non-Van Neumann modeling systems is important.  The ability of human controllers to make inputs to either the analog or digital models based on intuition or just trying to crash the system would also be valuable. This might be interesting enough for someone to write a grant for funding the initial phases.  What about Bob? :-)

As the Dutch said while fighting the Spanish: "It is not necessary to have hope in order to persevere."