Modeling Network Congestion under Demand Uncertainty Using Wardrop Principles

What the math gives to ML

The paper provides a constructive robust-bilevel viewpoint for evaluating congestion when demand is uncertain: an adversary chooses a demand realization while users respond according to a Wardrop equilibrium or a system-optimal flow. The most transferable ingredient is the budgeted uncertainty set, which produces structured worst-case stress tests rather than independent worst-case perturbations of every demand coordinate. This can be used to train mixture-of-experts routers or token dispatchers against concentrated but bounded shifts in token-type demand, with congestion represented by expert load, latency, overflow, or dropped-token penalties. The strongest initial implementation is a differentiable inner load-balancing equilibrium combined with a budgeted outer adversarial objective.