ChannelEstimation / pace_revision /docs /EXPERIMENT_MATRIX.md
TheHassanSaud's picture
Upload folder using huggingface_hub
c1e97d9 verified
|
Raw History Blame Contribute Delete
2.31 kB

Experiment matrix

The machine-readable list is generated by pace plan; it is copied into each initialized study. IDs encode experiment, recipe, architecture, construction, anchor/variant and training seed. A recipe's datasets are generated once and reused across compatible jobs.

ID Implementation Full fits
E01 Compact D, A-raw, A-res; seeds 41001–41020 60
E02 ReEsNet, Channelformer, FortiTran-M, HELENA; first ten seeds 40
E03 A-res with LS/LMMSE/sham/PH anchors; zero reuses D 40
E04 CNN16 and wide hybrid, D/A-res; Phase 2.1 amendment 40
E05 Five-plane information-preserving pair 20
E06 Source-normalized input/target pair 20
E07 Broad support/source prior; broad/isotropic; Kf2/Kt1; gamma0.1; gamma10 50
E08 Reverse shift: source fd100, test fd100/fd10 20
E09 Target retraining, frozen source BEM 20
E10 Broad-source training 20
E11 Comb6 and shifted pilot-symbol pair 40
E12 30-kHz numerology 20
E13 WAVE-source D/A-res/FortiTran/HELENA 40
E14 D trained on 109,000 source labels 10
R FortiTran-M / oracle-metadata AdaFortiTran-L; three seeds 6

Main total 440; R adds six. Development tuning adds 99 fits, resource gates 30 short fits, adaptation 40 short fits. Classical sweeps, oracles, predictions and receiver runs add computation but no neural training fits. No independent training is needed for zero anchor or fixed half mixture.

Main source: TDL-C, 100 ns, maximum Doppler 10 Hz. Main test conditions: C/100/{10,100,200,500} Hz; A/100/10; C/300/10; A/300/100. SNR 0,5,10,15,20,25 dB. Waveform receiver stress adds 1,000 Hz. Reverse and 30-kHz recipes have their separately declared target sets.

Source selection and test generation are separate CLI operations. Gate calibration is source-only. Target adaptation uses nested 128/1024 fitting labels plus 2,048 separate validation labels; both costs must be reported. Target oracles fit separate target TRAIN, never TEST.

The reference R is an independent published-setting reproduction using the released architecture and our canonical waveform engine. It does not claim exact historical MATLAB Figure 3(a) replication. Its sampled broad-domain coverage is not unseen-domain robustness.