DiLLSUE: a differentiable GPU solver for link-based logit stochastic user equilibrium
For transportation researchers and practitioners, DiLLSUE provides the first GPU implementation of link-based logit SUE with convergence guarantees, eliminating the need for route enumeration and training data.
DiLLSUE is a differentiable GPU solver for link-based logit stochastic user equilibrium that requires no route enumeration, training data, or network-specific tuning. It achieves 0.33% error on Sioux Falls, outperforming published trained surrogates, with GPU and CPU results agreeing to within 10^{-4}%.
Logit stochastic user equilibrium (SUE) captures imperfect route cost perceptions, but its practical solvers rely on route enumeration, which becomes prohibitive on large networks, or on link-based heuristics without convergence guarantees. We develop DiLLSUE, a differentiable solver for the link-based logit SUE that requires no route enumeration, no training data, and no network-specific tuning. Its inner loading algorithm batches all destinations into fixed-shape tensor operations, yielding, to our knowledge, the first GPU implementation of link-based logit SUE. Four outer solvers -- successive averages, self-regulating averaging, Anderson mixing, and implicit-function-theorem Newton -- are benchmarked under identical settings, and a family of acyclicity filters provides fast approximations with a quantified speed-accuracy trade-off. On five standard benchmark networks, GPU and CPU executions agree to within $10^{-4}$\% mean absolute percentage error, and the computed equilibrium converges monotonically to the independently computed Wardrop equilibrium, reaching 0.33\% error on Sioux Falls, where the zero-training solver is more accurate than published trained surrogates.