Finite-Time Convergence Rates in Stochastic Stackelberg Games with Smooth Algorithmic Agents

Authors: Eric Frankel, Kshitij Kulkarni, Dmitriy Drusvyatskiy, Sewoong Oh, Lillian J. Ratliff

ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 5. Empirical Vignettes In this section, we present some empirical results targeting some of the technical assumptions. These results elucidate the effect of the decision-maker s algorithm design space on the shape of the agent game and convergence behavior. A broader set of experiments is contained in Appendix D. D. Numerical Experiments Code is available at https://github.com/Sewoong Lab/stoch-stackelberg. In this section, we present numerical examples that are aimed at exploring the limits of the theory, and numerical examples that explore applying the theory to semi-synthetic real-world simulations. In particular, we use real-world data to create semi-synthetic experiments for price setting in ride-share markets.
Researcher Affiliation Academia 1University of Washington, Seattle 2University of California, Berkeley. Correspondence to: Eric Frankel <EMAIL>.
Pseudocode Yes Algorithm 1 Epoch-Based Drift-to-Noise Control Algorithm 2 Geometric Decay Schedule Algorithm 3 Projected Stochastic Gradient Play Algorithm 4 Epoch-Based Algorithm Framework for Stochastic Stackelberg Games
Open Source Code Yes D. Numerical Experiments Code is available at https://github.com/Sewoong Lab/stoch-stackelberg.
Open Datasets Yes D.3. Quadratic Ride-Sharing Game: Semi-Synthetic Simulations We consider an example from ride-sharing markets. Demand signals may be used to create more efficient ride share markets without reducing individual revenue streams by enabling information-limited firms to recover latent demand. Using an analogous set up as in Narang et al. (2022), we explore semi-synthetic competition between two ride-share platforms seeking to maximize their revenue given that the demand they experience is influenced by their own prices as well as their competitors. The data we use is from a prior Kaggle competition.4 4Data is publicly available: https://www.kaggle.com/datasets/brllrb/uber-and-lyft-dataset-boston-ma
Dataset Splits No The data we use is from a prior Kaggle competition.4... It uses "semi-synthetic simulations" but doesn't specify how the Kaggle data was split into train/test/validation for their specific experiments. It's more about "game formulation" using data characteristics. The "simulations" use "random quadratic game instance" and "synthetic Kelly auction".
Hardware Specification No No specific hardware details are mentioned in the paper.
Software Dependencies No No specific software versions or dependencies are mentioned in the paper.
Experiment Setup Yes 5. Empirical Vignettes In this section, we present some empirical results targeting some of the technical assumptions. These results elucidate the effect of the decision-maker s algorithm design space on the shape of the agent game and convergence behavior. A broader set of experiments is contained in Appendix D. D.1. Effect of Choosing τ without a priori Knowledge of Agent Game Parameters ...we explore the choice of τ on the tracking error of the agents equilibrium relative to the appropriate equilibrium... ...the decision-maker is deploying actions to with the objective of tracking a desired equilibrium. The decision-maker may not know the agent game-related hyperparameters a priori and thus may not be able to set τ optimally as noted. D.3.1. Estimation of Price Elasticities ...if a decision-maker aims to estimate each of the Ai s i.e., the price elasticities of players then they can run online least squares where in each round t they first query the environment and observe zt,i = ξt,i + Aixt where Ai = Ai,i Ai, i . Then, they perturb the prices with actions ui for each player, and observe qt,i = Ai(xt + ut,i) + ξ t,i. With these two queries, the decision maker updates their estimate of the price elasticities as follows: ˆAt+1,i = ˆAt,i + νt(qt,i zt,i ˆAt,iut,i)u t,i, where νt is the step size. Theorem 4.4 ... using step-size η α 4L2 ℓ(1+L2eq)... epoch length τ O 1 (1 ρ2) log 2 R2 ϵτ + c2σ2 a ρ2.