Collapsing Bandits

Code repo for "Collapsing Bandits" (NeurIPS 2020) simulation environment. Main paper is here and full paper with appendix is here.

@article{mate2020collapsing,
  title={Collapsing Bandits and Their Application to Public Health Intervention},
  author={Mate, Aditya and Killian, Jackson and Xu, Haifeng and Perrault, Andrew and Tambe, Milind},
  journal={Advances in Neural Information Processing Systems},
  volume={33},
  year={2020}
}

Execution

Example Usage: python3 adherence_simulation.py -N 10 -n 50 -k 5 -s 0 -ws 0 -ls 0 -l 180 -d simulated

• -N: Number of trials to average over
• -n: Number of processes
• -k: Number of active actions allowed each day
• -s, ws, ls: Seeds for random generator streams.
• -L: Length of simulation (default is 180 rounds, to simulate 6 months of TB adherence data)
• -d: Dataset to use (real/simulated). Note that real TB data is not publicly available, so has been removed from codebase, so the code can only run in the "-d simulated" mode.

Note: The above command will run the simulation for the faster policies only since the baselines take lot of time to run. To run other baselines add a -pc flag to run a specific policy, with parameters as exlplained below:

For a specific policy python3 adherence_simulation.py -N 10 -n 50 -k 5 -s 0 -ws 0 -ls 0 -l 180 -pc 5

• -pc 0: No actions
• -pc 1: All actions are active
• -pc 2: Random actions
• -pc 3: Myopic policy
• -pc 5: Threshold Whittle
• -pc 10: Qian et al.
• -pc 14: Oracle