The trouble here is that Roughgarden’s model seems to be based on speculation on the selfish actor theory, but ignores real world phenomena. For example, he references Braess Paradox in his paper, and yet under real world conditions, the paradox disappears when under high load: https://supernet.isenberg.umass.edu/articles/NegationoftheBraessParadox.pdf Which is strangely consistent with more accurate simulation / agent based models of behavior that also deviates from the prescribed model of how selfish routing works. http://dl.acm.org/citation.cfm?id=2570361
As far as I can gather, this discrepancy is because the selfish routing models depend on a simplification/assumption of the agent’s goals and ignores the the more complex signaling games and memory systems that stem out of the real-world dynamics of game theory. Unless I’m misunderstanding, this is basically shooting down a game-theoretic strawmanificaiton of the real world?
To be more specific, the assumptions can be seen in the formulation of the definition of Wardrops’s Equillibrium, which comes as a result of this framing
“Naturally we expect selfish traffic to attempt to minimize cost” (per his paper on the subject) — Primarily in expectation that this is selfish traffic.
It’s the same kind of assumption that makes for comparisons of traffic to the prisoner’s dilemma, were it is much more apt to consider it some form of a Public goods game under constraints of resource density and temporal variation factors which lead to a non-steady state equilibrium effect as energy (congestion) is added to the system. Different game model = different outcomes.
TL;DR: Selfish Routing argument, while sounding good on the surface, is not an adequate description of the complex problem space of traffic.
