1_simple_regret dict | 2_cumulative_regret dict | 3_info_gain dict | 4_gap_closure dict |
|---|---|---|---|
{
"claim": "1_simple_regret",
"verdict": "PASS",
"predicate": "0.2 < avg_growth_rate_k < 3.0 and R2(log(regret) vs d, mean curve) > 0.7",
"avg_regret_by_d": {
"2": 0.0015900002160488523,
"3": 0.01949763251203758,
"4": 0.16273679919352826
},
"ratio_d3_d2": 12.262660291008741,
"ratio_d4_d3": 8.3... | {
"claim": "2_cumulative_regret",
"verdict": "PASS",
"predicate": "0.3 < avg_power_law_exponent < 0.9",
"avg_power_law_exponent": 0.5536016115136881,
"exponent_ci95_halfwidth": 0.02504889999854189,
"avg_r2": 0.9701823186059776,
"T_vals": [
80,
120,
180,
250
],
"per_seed": {
"0": {
... | {
"claim": "3_info_gain",
"verdict": "PASS",
"predicate": "0.05 < avg_log_log_slope < 0.8 (lower bound guards against the degenerate all-zero bug)",
"avg_log_log_slope": 0.19493026723185172,
"slope_ci95_halfwidth": 0.0009389135397642761,
"avg_r2": 0.9976570470639831,
"theoretical_bound_slope_same_grid": 0... | {
"claim": "4_gap_closure",
"verdict": "PASS",
"predicate": "-1.0 < avg_improved_power < 2.5 (prior/un-improved exponent is d+1=3)",
"avg_improved_power": 0.04571713958725035,
"improved_power_ci95_halfwidth": 0.009819964016643886,
"avg_r2_full_fit": 0.9998545294069209,
"note": "Improved power well below p... |
Reproduction bundle: Tighter Regret Lower Bound for GP Bandits with SE Kernel (ICML 2026 repro)
Bundle for the logbook Space algorise/repro-tighter-regret-lower-bound-for-gaussian-process-bandits-with-squared-exponential-kernel-in.
repro/run_experiments.py- CPU-only experiment suite (30 seeds/claim, 95% CIs, pre-stated two-sided pass predicates). ~12.7 min on a shared CPU box.repro/make_figures.py- regenerates the embedded claim-page figures from the results JSON.results/tighter_regret_gp.json- full per-seed results (all 4 claims PASS).
Rerun: pip install numpy scipy matplotlib then python repro/run_experiments.py.
- Downloads last month
- 38