IMHO, fully random assemblies work best in situations where nearly all tiles can be placed next to each other, such as +oriental or +forest. Unfortunately, these use-cases tend to be limited to natural areas. As soon as you introduce an element such as roads, tunnels or human-built areas which have to look planned in some way, it's a bit more difficult. The RMA algorithm will, impressively, build a map if it has the right pieces. But it's nearly impossible to get the maps to come out looking like a real place.
If you're working with tunnels or streets, you might consider using fixed layouts alongside tilesets. In this way, you can lay out a street grid, for instance, and then have the RMA insert random buildings at the appropriate slots. Because you can create as many different assemblies as you like, and they take very little time to create, you can easily set up several different street grids to prevent the same exact layout from occurring every time.
You can actually get more random-ness this way, I think, than by going with the fully random assemblies. This is because the algorithm tends to stumble upon similar solutions given a recurring configuration of tiles. For instance, although you can use lots of "filler" tiles to fill in gaps and ensure a layout is found, you tend to end up with a map packed with fillers and too much space (see +village).
That said, I'd love to see some experiments with tile size and the RMA, using simple blocks. How do layout results differ if I use all 2x2 and 2x4 tiles? What if I have a 2x1 a 2x2 and a 2x3 title? I suspect we can come learn some lessons about what works best to produce different layout results if we test like this.
You MIGHT be able to make some kind of no-clipping, invisible-but-shadow-receiving brush to fudge this gap...
Look for the Lightclip texture. Check out the +alienb map to see how the tunnels are stitched together (or the +b maps).