Branch-point Ensemble Timing Grid

Generated from outputs/tables/branchpoint_ensemble_selection_timing_grid.csv and outputs/tables/branchpoint_ensemble_selection_timing_blocks.csv.

What exists now

Timing grid rows: 4375.
Timing block rows: 120.
Dimensions: binary and multistate feature matrices; prefix and greedy p-adic ensembles; feature subset sizes s=5..10; selected ensemble sizes up to m=200; K=200 candidate libraries; 10 repeats.
The old score curve existed before this timing pass; the runtime columns were added by instrumenting candidate generation and prefix/greedy selection replay.
mean_total_fixed_library_elapsed_seconds times the full K=200 candidate generation plus selection. mean_estimated_total_k_equals_m_elapsed_seconds prorates generation to K=m and adds measured selection time.

dataset_label	method	subset_size_s	count	min	max
Binary	greedy	5	199	1	199
Binary	greedy	6	197	1	197
Binary	greedy	7	186	1	186
Binary	greedy	8	173	1	173
Binary	greedy	9	136	1	136
Binary	greedy	10	93	1	93
Binary	prefix	5	200	1	200
Binary	prefix	6	200	1	200
Binary	prefix	7	200	1	200
Binary	prefix	8	200	1	200
Binary	prefix	9	200	1	200
Binary	prefix	10	200	1	200
Multistate	greedy	5	198	1	198
Multistate	greedy	6	194	1	194
Multistate	greedy	7	188	1	188
Multistate	greedy	8	176	1	176
Multistate	greedy	9	141	1	141
Multistate	greedy	10	94	1	94
Multistate	prefix	5	200	1	200
Multistate	prefix	6	200	1	200
Multistate	prefix	7	200	1	200
Multistate	prefix	8	200	1	200
Multistate	prefix	9	200	1	200
Multistate	prefix	10	200	1	200

dataset_label	method	subset_size_s	ensemble_size_m	mean_padic_loss	se_padic_loss	mean_total_fixed_library_elapsed_seconds	mean_estimated_total_k_equals_m_elapsed_seconds
Multistate	greedy	8	14	0.003	0.0024	4.9051	0.5781
Binary	greedy	9	5	0.0083	0.0029	1.4816	0.1096
Binary	prefix	7	94	0.0383	0.0024	5.7849	2.7249
Multistate	prefix	5	194	0.046	0.0011	7.8844	7.6486

dataset_label	method	subset_size_s	ensemble_size_m	mean_padic_loss_summary	se_padic_loss_summary	mean_exact_branch_accuracy	mean_active_parameters_used	mean_total_fixed_library_elapsed_seconds	mean_estimated_total_k_equals_m_elapsed_seconds
Binary	greedy	9	4	0.0093	0.0026	0.8667	9.3917	1.4668	0.0878
Binary	prefix	7	94	0.0383	0.0024	0.6417	249.183	5.7849	2.7249
Multistate	greedy	8	5	0.005	0.0026	0.9417	26.0917	4.7458	0.2094
Multistate	prefix	5	194	0.046	0.0011	0.5167	1022.07	7.8844	7.6486

dataset_label	subset_size_s	repeats	mean_total_seconds	mean_candidate_generation_seconds	mean_prefix_selection_seconds	mean_greedy_selection_seconds	mean_candidate_rows
Binary	5	10	8.2395	5.8944	0.0239	2.2648	2201.4
Binary	6	10	8.9432	6.689	0.0241	2.1761	2061.2
Binary	7	10	7.7946	5.7736	0.0239	1.9484	1818.6
Binary	8	10	5.1415	3.5218	0.0237	1.5557	1468.6
Binary	9	10	2.4576	1.4071	0.024	0.999	882.6
Binary	10	10	0.7345	0.3334	0.0238	0.3633	264.8
Multistate	5	10	10.2045	7.8614	0.0238	2.2614	2263.1
Multistate	6	10	11.404	9.2531	0.0237	2.0723	2121.6
Multistate	7	10	9.7999	7.8286	0.0238	1.8967	1927.7
Multistate	8	10	6.2826	4.6527	0.0235	1.5627	1603.1
Multistate	9	10	2.8397	1.7307	0.0237	1.0547	1017
Multistate	10	10	0.8204	0.3886	0.0236	0.393	317.7

Greedy score curves: outputs/figures/branchpoint_ensemble_timing_grid_greedy_loss_curves.png
Greedy estimated K=m runtime curves: outputs/figures/branchpoint_ensemble_timing_grid_greedy_time_curves.png

Greedy selection dominates prefix on loss and reaches its useful region with small m; the one-SE choices are m=4 for binary and m=5 for multistate.
Increasing s helps through about s=8 or s=9, but s=10 is worse even though it is faster because fewer candidate through-point fits are feasible.
Candidate generation is the main cost at fixed K=200; replaying prefix selection is negligible and greedy selection is usually around one to two seconds per dataset/subset/repeat.