Search arXiv⌕ Search
Home/Archive/arXiv

arXiv · 2610.04337

FLAT: Smoothing the Rugged Landscape for Learnable, Sample-Efficient Traffic Calibration

Haopeng Deng·Shuo He·Dayuan Wang
View original source ↗Read source document ↗Export metadata ↓

Abstract

Calibrating microscopic traffic models for digital twins is an expensive black-box optimization problem: tuning car-following and lane-changing parameters requires a full simulation run, affording only a tight budget per recalibration window. Matching raw trajectories yields a rugged objective that sparse surrogates cannot learn, reducing sequential acquisition to near-random probing. We present FLAT, which couples what to optimize with where to sample next. An eight-dimensional behavioral fingerprint smooths the parameter-error landscape, making the objective learnable from a few dozen samples; annealed lower-confidence-bound (LCB) acquisition then spends each remaining run where it most reduces error. The surrogate, interchangeable among a Gaussian process (GP), random forest (RF), or multi-layer-perceptron (MLP) ensemble, plugs into the same LCB loop. Across six heterogeneous real-world scenes, FLAT-GP achieves the lowest scene-averaged behavioral error, winning 6/6 scenes against SPSA, GA, and CMA-ES and 5/6 against TPE under the matched budget. Some baselines need up to 4.4 times more simulations to match. Ablations show objective choice shifts final behavioral error by 81% on average, removing sequential LCB raises the six-scene mean by 20%, and surrogate choice shifts it by at most 4.2%, confirming gains trace to objective geometry and sequential allocation rather than surrogate capacity.

Explore related subjects

physics.soc-phcs.CEcs.LG

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Haopeng Deng, Shuo He, Dayuan Wang. 2026-10-03. FLAT: Smoothing the Rugged Landscape for Learnable, Sample-Efficient Traffic Calibration. https://arxiv.org/abs/2610.04337

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

Functional Dismantling of Network Relaxation through Slow-Branch Susceptibility

Robustness of relaxation on asymmetric networks is not determined by connectivity alone, because the slow collective mode can be complex and may change its spectral identity under adaptive damage. We introduce a slow-branch susceptibility framework for functional dismantling of network relaxation. Starting from the projected relaxation dynamics, we show that the relevant robustness observable is the real part of the selected nonzero Laplacian branch, which controls the long-time decay of the nonstationary sector. Node deletion is then treated as a dimension-changing compression of the operator, leading to a modal susceptibility (MS) score that estimates the first-order reduction of the branch-tracked relaxation rate from the biorthogonal support of the slow mode. In the reciprocal limit, the same construction reduces to the weighted Fiedler sector, placing directed and weighted-undirected networks within a common spectral-response formulation. Tests on synthetic and real-world networks show that MS identifies vulnerability patterns that differ from standard centrality-based attacks and edge-level spectral proxies. These results resolve a modal-selection ambiguity in non-Hermitian robustness analysis and provide a spectral basis for functional dismantling in asymmetric networks.

physics.soc-ph↗
arXiv2026

Hierarchical Synchronization and Distortion Scaling in Social Media Networks: A Fractal-Like Topology Theory

The rapid proliferation of social media as a dominant channel for information dissemination has intensified concerns over systemic information distortion, whereby content is progressively altered through successive layers of transmission. While prior studies have explored such distortion qualitatively, the quantitative interplay between propagation topology and stochastic cognitive perturbations remains insufficiently understood. In this work, we propose a novel fractal-inspired directed hierarchical network model to capture the structural patterns of propagation, and introduce a Noise-Frustrated Hegselmann-Krause (NFHK) framework to model opinion dynamics under noise. Analytical results, supported by group and graph theory, reveal that noise accumulation leads to increasing opinion distortion and the emergence of intra-layer synchronization. Multi-agent simulations confirm these effects, showing that noise intensity shapes both convergence rates and weak intra-layer clustering. Empirical validation using a representative retweet cascade demonstrates that the proposed model reproduces real-world distortion patterns and synchronization behaviors, even without direct links. This work uncovers a unified mechanism for information distortion in digital platforms and offers topology-aware insights for public opinion governance and platform regulation.

physics.soc-ph↗
arXiv2026

Human-AI Interaction Mediated by Courtship Language Culture: Computational Statistics and Probabilistic Modeling

Large language models are entering living cultural practices, including intimate ones. We study a collegiate courtship tradition at Dartmouth College called flitzing, the tradition of sending rhymed flirtatious emails to ask peers out, as a measurable microcosm of a culture absorbing artificial intelligence. We combine four sources of evidence: human evaluation surveys in two waves (2025 and 2026); information retrieval over 33 human-written and 5,610 AI-generated flitzes spanning 17 sampling temperatures; evaluation of the same survey by three generative language models; and simulation of a coupled multi-species bipartite spin glass model of people and texts. Four main findings emerge. First, sampling temperature induces a phase transition in diversity: below the transition, AI-generated flitzes are more homogeneous than the human, while above it they decohere, and human baselines of diversity remain stable across years, indicating persistent collective structure in human creativity. Second, human judges conflate perceived quality with human authorship, whereas AI judges detect provenance almost perfectly yet still rate human-written flitzes higher. Third, the transition window, with temperature between 1.25 and 1.38, is consistent across biased survey data, unbiased text statistics, and the mathematical model, which also reproduces the observed clustering. Fourth, detection ability generalizes beyond the local culture, with Dartmouth and non-Dartmouth respondents performing comparably, and strengthens over time, with 2026 respondents above chance at every temperature. We conclude that AI participation in a tight cultural genre is regime dependent: conformist and homogenizing below criticality, detectable and disruptive above it.

physics.soc-ph↗