This repository contains the experimental code for the NeurIPS 2025 paper:

"Design-based Bandits Under Network Interference: Trade-off Between Regret and Statistical Inference".

We provide the experimental scripts:

• Experiments in the main body of the paper:

  • exp_101units_adv.py: Simulates a 101-unit star network for multi-armed bandit (MAB) experiments under network interference.

• Experiments in Appendix E (Additional Experimental Results):

  • exp_singleunit_adv.py: Instance 1
  • exp_6unit_adv_222.py: Instance 2
  • exp_10unit_adv_145.py: Instance 3
  • exp_10unit_adv_1333.py: Instance 4

All scripts evaluate the performance of three exploration-exploitation strategies:

• Uniform: Fully exploratory approach.
• Standard EXP3: Traditional EXP3 method for regret minimization.
• EXP3-N-CS: Our proposed method balancing regret minimization and statistical inference.

Execute the following commands to run the experiments:

# Example:

# Run the 101-unit star network experiment
python exp_101units_adv.py

# Run the 5-arm multi-armed bandit experiment
python exp_singleunit_adv.py

Take exp_101units_adv.py as an example:

• Simulates a 101-unit star network with 1 central node and 5 outer clusters.
• Each cluster follows a shared action assignment.
• Implements exposure mapping to define unit responses based on neighboring actions.
• Compares Uniform, Standard EXP3, and EXP3-N-CS under an adversarial reward schedule.

• Cumulative Regret: Lower values indicate better exploration-exploitation balance.
• CS Width: Smaller widths indicate more precise statistical inference.
• ATE Estimation Error: Smaller error means better inference performance.

The EXP3-N-CS method demonstrates a trade-off between minimizing regret and improving statistical inference, outperforming Standard EXP3 in inference accuracy while still maintaining competitive regret.

• Modify the network topology in exp_101units_adv.py by adjusting the adjacency matrix in section (A).
• Adjust the horizon T and number of replicates N_exp in both scripts to test different settings.
• Change the exploration parameter delta_t in EXP3-N-CS to observe different trade-offs.

After running the scripts, you will see:

1. Cumulative Regret comparison across methods.
2. CS Width evolution for the most challenging inference pairs.
3. ATE Estimation Error distribution.

This code is for academic and research purposes.