Information Design for Congestion Games with Unknown Demand

Authors: Svenja M. Griesbach, Martin Hoefer, Max Klimm, Tim Koglin

AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Finally, we conduct a computational study that tests this algorithm on real-world instances. ... We conducted a computational study exhibiting that the number of different supports used in a Wardrop equilibrium is small on real-life instances. Therefore, our LP-based algorithm can be implemented in reasonable time in practice. ... We considered non-atomic network congestion games with affine costs and uncertain demand on real-world networks for a single commodity and two possible states of nature Θ = {θ1,θ2}. Table 1 shows the six different networks we examined. ... We performed 40 simulations for each network with varying (s,t)-pairs. ... Table 2 shows averaged results on the properties of Ai. ... Table 3: Performance of full information revelation (FI), no-signaling (NO), and the optimal signaling scheme (OPT) averaged over 40 instances for each network with µ θ2 = 0.5.
Researcher Affiliation Academia 1Department of Mathematics, TU Berlin, Germany 2Institute for Computer Science, Goethe University Frankfurt, Germany {griesbach, klimm}@math.tu-berlin.de, {mhoefer, koglin}@em.uni-frankfurt.de
Pseudocode No The paper describes algorithms (FPTAS, LP-based algorithm) but does not provide them in structured pseudocode or algorithm blocks.
Open Source Code Yes All source codes and data sets are available on Git Hub (Griesbach et al. 2023a).
Open Datasets Yes The network data was obtained from the Transportation Networks for Research Core Team (2022). ... Transportation Networks for Research Core Team. 2022. Transportation Networks for Research. https://github.com/bstabler/Transportation Networks. Accessed: 2022-01-14.
Dataset Splits No The paper describes simulation setups on network instances (e.g., "40 simulations for each network with varying (s,t)-pairs") but does not mention standard training/validation/test splits commonly found in machine learning contexts.
Hardware Specification Yes Experiments were performed on an Intel Core i5 based computer at 3.47 GHz with 8 GB RAM operating on Ubuntu 22.04.1 LTS.
Software Dependencies Yes We used the built-in solver of the Sci Py package (v1.8.1). The flow assignments were computed by an implementation of the conjugate Frank-Wolfe algorithm... in Python (v3.10.6) based on the code of Bettini (2022).
Experiment Setup Yes The alternative demand dθ1 was defined relative to dθ2, i.e., dθ1 = ρ dθ2 for some ρ [0,1]. In the following, we show results for ρ = 0.2. We performed 40 simulations for each network with varying (s,t)-pairs. For each simulation, the (s,t)-pair was drawn uniformly at random from the set of zones such that s /= t and no pair was chosen more than once. The flow assignments were computed by an implementation of the conjugate Frank-Wolfe algorithm (Frank and Wolfe 1956; Daneva and Lindberg 2003) in Python (v3.10.6) based on the code of Bettini (2022). More information on used libraries and parameters is provided in the full version (Griesbach et al. 2023b, Appendix D).