I don't rule out a very good programmer making intelligent rules of abstraction for getting the branching problem under control, to quickly run simulations. But I don't think it's ever been done yet (for crpgs).
Not necessary. eRe4s3r and me (and others) have independent from each other already identified some of the glaring combat imbalances to begin with.
I believe the first half my post was saying a lot more can be done for removing some obviously bad choices. But once you try analyzing future turns, it gets very hard computationally.
A game with a far smaller strategy space like limit-hold'em is still very difficult for computers to analyze, and challenge decent human players. It's been done though, it just took a very long time for computers / programmers to get there. It might be a little easier than I think, b/c unlike poker there aren't (or don't need to be) hidden state variables (cards face down). It's harder for the computer, for instance, when it may not know the player's resistances to certain damages until it attacks. This introduces more branching earlier.
Proceeding under assumptions of what an average Joe would do is how nearly all computer game "AIs" are programmed. I put AI in quotes, because its not an AI, just a human telling it what a computer should consider and how it should react, rather than the computer considering the costs / benefits itself.
Human-scripted AIs are really hit and miss, and often fairly predictable. For instance, you may tell the computer to take out the healer, and after a few fights the player realizes the AI is always going to go after a healer, so turns his healer into a tank, negating what the programmer thought would create challenge. The ability of a scripted-AI to adapt is limited by the script which is by necessity fairly basic, and it will generally play the same "hand" the same way, every time, making it predictable too.