Spectral bandits for SOR ω

Learners pick a relaxation parameter ω each round. Feedback is SOR iteration count on a Poisson (5-point Laplacian) system, or a precomputed full-info loss tensor.

Setup

conda activate banditos-clean   # or: pip install -e .
pip install -e .
python -m pytest

Configs are Python files. A config builds already-parameterized learners (the graph lives on the spectral algo) and a data source, then calls run(algos, data). Sweeps are ordinary loops in the config.

python scripts/run.py configs/live_sor.py
python scripts/run.py configs/replay_pitzdaily.py
python scripts/run.py configs/sweep_alpha.py
python scripts/run.py configs/sweep_k.py

Optional Weights & Biases: install wandb yourself and set WANDB_PROJECT, or pass wandb_project=... into run(). It is not a package dependency.

Paper figures

Regret vs timestep (live SOR):

python scripts/run.py configs/live_sor.py

Same plot from archived losses (data/fullinfo/pitzDaily.npy):

python scripts/run.py configs/replay_pitzdaily.py

Each run() writes plots/runs/<stamp>_<name>/:

• metrics.csv — per-timestep mean cost / cumulative loss
• summary.csv — final totals
• run.txt — names, HPs, graph, data, seed, T, K, trials
• arrays.npz — cost and action-probability arrays

CSV losses work the same way (load_losses("losses.csv", similarity_kind="chain"), T×K, one trial). Then pass the bundle into run() like the replay config.

Arm probabilities vs time are written next to the regret plot. Replot from a saved .npz:

python scripts/plot_regret_t.py plots/runs/<run>/arrays.npz
python scripts/plot_arm_probs.py plots/runs/<run>/arrays.npz

Regret vs number of configs (K):

python scripts/run.py configs/sweep_k.py
python scripts/plot_regret_k.py plots/sweeps/<stamp>_k/results.csv

Basic hyperparameter sweep (offline α)

python scripts/run.py configs/sweep_alpha.py

Layout

src/ltr/learners/     Tsallis-INF, Exp3-Spectral, Tsallis-Spectral
src/ltr/solvers/      SOR
src/ltr/data/         load .npy / .npz / .csv losses
src/ltr/graphs/       chain / ring / double_chain + load K×K
src/ltr/plot/         the three figures
src/ltr/experiment.py run(algos, data)
configs/              Python configs (live, replay, sweeps)
scripts/run.py        python scripts/run.py configs/foo.py

Old experiment dumps (pitzDailyLarge, sweeps, etc.) are in ../learning-to-relax-archive/, not in this tree.