Performative Prediction on Games and Mechanism Design

Repository for "Performative Prediction on Games and Mechanism Design".

AISTATS 2025. Góis, A., Mofakhami, M., Santos, F. P., Gidel, G., & Lacoste-Julien, S.

Instructions

Simulation for RRM with scale-free networks (Figure 4):

• python plot_rrm_scalefree.py --save_path rrm

Heatmaps for the anarchic setting with \tau=0 (Figure 5):

• python plot_alpha_heatmap.py --graph full
• python plot_alpha_heatmap.py --graph scale-free

Trust oscillation (Figure 6):

• python plot_trust_oscillation.py --discount_rate med

Tradeoffs between accuracy and welfare (Figure 7):

• python plot_tradeoff_thresholds.py

Plot to compare architectures (Figure 8):

• python train_all_archs.py --seed 0 --stats_path crd_archs_stats --loss group
  • repeat for seeds 1, 2, 3
• python plot_architectures_avgs.py --read_path crd_archs_stats --loss group

Histograms for RRM with scale-free networks in appendix (Figure 12):

• python plot_rrm_scalefree.py --plot_single_pop -n 20
  • repeat for n=30, 50

Ablation of gradient components in appendix (Figure 15):

• python train.py --architecture gnn+mlp --seed 0 --epochs 200 --topology scale-free --loss individual -lr 1e-4 --save_stats --use_custom_grad --block_prev_trust --stats_path crd_stats_grad-blockprevtrust/
  • repeat for seeds 1, 2, 3, 4
• python train.py --architecture gnn+mlp --seed 0 --epochs 200 --topology scale-free --loss individual -lr 1e-4 --save_stats --use_custom_grad --block_trust_grad --stats_path crd_stats_grad-blocktrust/
  • repeat for seeds 1, 2, 3, 4
• python train.py --architecture gnn+mlp --seed 0 --epochs 200 --topology scale-free --loss individual -lr 1e-4 --save_stats --use_custom_grad --stats_path crd_stats_grad-full/
  • repeat for seeds 1, 2, 3, 4
• python plot_ablation.py