PPK in C++: 330 seconds, R-precision: .394. That's with a covariance floor of 0.00001.
Monte Carlo KL-d in C++, 2000 samples: 2 hours, R-precision .439. Best so far.
KL-EMD in C++: 20 minutes, R-precision .301.
On uspop1000:
PPK in C++: 45 minutes, R-precision: .440
KL-emd: +8 hours(?), R-precision: .396
Finally, a few PPK kernels built from models with different numbers of Gaussians, on uspop1000:
50: .4402
30: .4396
20: .4382
Posted by madadam at July 24, 2006 11:36 AM