Code for the experiments in Bandits with Untrusted Change-Time Forecasts (submitted to ACM Transactions on Algorithms, 2026).
FA-UCB is a UCB-style algorithm for piecewise-stationary multi-armed bandits that receives untrusted predictions of when the reward distributions change. Each prediction is verified by a paired statistical probe (fresh samples taken just before and just after the predicted time); the learner restarts only on a confirmed change. A geometric distrust gate bounds the probing cost of junk predictions logarithmically.
| Script | Paper artifact | Runtime* |
|---|---|---|
exp_main.py | horizon-scaling figure (M-UCB series) + adversarial figure | ~25 min |
exp_glr_and_gate.py | horizon-scaling figure (GLR series, T ≤ 200k) + gate figure | ~50 min |
exp_glr_400k.py | horizon-scaling figure (GLR point at T = 400k) | ~45 min |
exp_gate.py | gate figure (overlapping-prediction schedule) | ~4 min |
exp_easy_regime.py | negative-scope table (easy regime) + asymmetry sweep | ~5 min |
exp_dense_blindspot.py | negative-scope table (dense blind-spot regime) | ~12 min |
make_figures.py | renders figures/*.pdf from the numbers reported by the scripts | seconds |
*Measured on a laptop CPU (Intel i5-1235U, single process). No GPU is used.
pip install -r requirements.txt
python exp_main.py # writes results/main.txt
python exp_glr_and_gate.py # writes results/glr_and_gate.txt
python exp_glr_400k.py # writes results/glr_400k.txt
python exp_easy_regime.py
python exp_dense_blindspot.py
python exp_gate.py
python make_figures.py # writes figures/*.pdf
All randomness is seeded (environment seeds and algorithm seeds are separate;
matched seeds give paired comparisons across algorithms). Each script prints
its summary table and writes it under results/.
- The GLR baseline follows the recipe of Besson, Kaufmann, Maillard and Seznec (arXiv:1902.01575) with three deviations for CPU feasibility, documented in the script header and in the paper.
- The theoretical analysis uses a horizon-aware exploration bonus; the experiments run standard UCB1. This is disclosed in the paper.
MIT (see LICENSE).