I think the weakness of R[xx] systems is that they're only working one move at a time. Any position suggests a number of strategies to a good player. Being stuck with a randomized system that tends towards plausible moves means that a move that builds in one direction could be weakened in the next round by a move that's tending towards another good strategy.

Could R[xx] be improved (in theory--I think the calculations would be unmanageable) by considering more than one move at a time? Is there some number of moves where you don't exactly have an R[xx] anymore because the program is more like simulating the way the real chess player thinks?