hardware-counters
advanced_nn_summary.txt
Advanced tabular NN techniques
==============================================================================
191 merged configurations, 21 features, 8 applications
inner search grid (random subsample of 12): s0.05/k32/w128, s0.2/k32/w256, s0.05/k16/w128, s0.2/k32/w128
Constant (median log-eta) median 1.0654 p90 1.5324 [0s]
Random Forest median 1.0548 p90 1.3887 [7s]
GBM quantile(0.5) median 1.0726 p90 1.4018 [6s]
PLR ensemble (fixed defaults) median 1.1854 p90 1.6744 [122s]
PLR ensemble (nested select) median 1.1646 p90 1.6885 [114s]
inner picks: s0.05/k16/w128, s0.05/k16/w128, s0.2/k32/w256, s0.2/k32/w256, s0.05/k32/w128, s0.05/k32/w128, s0.2/k32/w256, s0.2/k32/w128
RealMLP-TD (pre-tuned) median 1.0829 p90 1.3537 [39s]
PLR + RF hybrid median 1.1127 p90 1.5165 [74s]
--- paired Wilcoxon vs constant baseline (n=191) ---
Random Forest wins 124/191 p=0.0002445 (better)
GBM quantile(0.5) wins 118/191 p=0.0004311 (WORSE)
PLR ensemble (fixed defaults) wins 57/191 p=2.965e-06 (WORSE)
PLR ensemble (nested select) wins 60/191 p=3.993e-06 (WORSE)
RealMLP-TD (pre-tuned) wins 83/191 p=0.9833 (WORSE)
PLR + RF hybrid wins 77/191 p=0.1226 (WORSE)
--- per-application median ---
Constant (median log-eta) Random Forest GBM quantile(0.5) PLR ensemble (fixed defaults) PLR ensemble (nested select) RealMLP-TD (pre-tuned) PLR + RF hybrid
app
comd 1.046 1.028 1.027 1.219 1.117 1.133 1.117
gromacs 1.051 1.007 1.015 1.187 1.167 1.008 1.082
hpcg 1.048 1.117 1.081 1.176 1.201 1.077 1.102
hpl 1.078 1.034 1.075 1.156 1.131 1.054 1.077
lulesh 1.175 1.048 1.110 1.095 1.075 1.045 1.072
minife 1.372 1.062 1.151 1.160 1.160 1.291 1.102
openfoam 1.068 1.075 1.106 1.156 1.279 1.092 1.128
stream 1.051 1.096 1.078 1.618 1.600 1.064 1.331