On the face of it, solution concepts seem like a useful tool. You’re making some assumptions when you talking about behavior, and solution concepts let you lock in those assumptions and then treat them formally.

The issue, of course, is that you take this enormous space of what could happen in an interaction, and choose to focus only on what your model says will happen. And most perversely, we seem to actively encourage models which are as specific as possible; the best models are those which promise unique solutions. This dimensionality reduction is a brittle process that leads to ridiculous, counter-intuitive, and empirically false results (like the Revenue Equivalence Theorem).

Essentially, current approaches are maximally specific, but one thing we know from AI is that you can produce much more accurate predictive models by being maximally general. Say you have some history about the way people are behaving. The least specific model that is consistent with this history, the one with maximum entropy, will almost certainly be the best predictor of actual behavior. My favorite demonstration of maximum entropy is Brian Ziebart’s video from his research on predicting the destination of drivers. As an analogy, consider that current solution concepts would just say “He’s going to the airport” and leave it at that.

We should strive for robust models, and you can’t get robustness without embracing uncertainty.