To do

• modify asd in other repo to have correct evidence

• check out gridded results

• compare evidence/ll of my solution, jake's solutions

  --ro 19.069070174197527 --ssq 2.5019481075308638 --delta 10.526593885061727 evidence=627341.397251 neg. log likelihood=3119100.34894

  --ro 17.238799208518522 --ssq 3.1579704282716046 --delta 23.578709766296299 evi=764016.10162078473

Ideas

• regress where foe moves towards eye, foe moves away from eye

Solutions

--ro 7.06184735 --ssq 56.27155504 --delta 0.12

--ro 8.84497231 --ssq 9.9393613 --delta 0.51312846

--ro 8.84497231 --ssq 9.9393613 --delta 0.5
    evi=-2629.9342857768866
    nll=7444.9439632018575

--ro 8.84497231 --ssq 9.93496026  --delta 0.51151012
    evi=-2629.9265325811289
    nll=7448.1781925010782

--ro 7.06184735 --ssq 56.27155504 --delta 0.12
    evidence=-700.661732086
    neg. log likelihood=1356.99041683

--ro 9.2329678 --ssq 1.64542619  --delta 4.98630877
    TNC max dude, starting from Ridge solution
    evidence=-11298.0465625
    neg. log likelihood=44957.8554228

Jake

Active-Set

--ro 10.27167 --ssq 1.468826 --delta 13.236098
--ro 11.0 --ssq 1.3 --delta 10.7
    evidence=2262.81097632
    neg. log likelihood=96013.442144

log these, then exp
(2.0, 20.0) -> 11
(1.0, 2*10e4) -> 1.3
(1.0, 2*10e4) -> 10.7

Notes

• chol too conservative? using LU logDet still allows calculation in cases where you're doing SVD trick
• n.b. ASD/RD is taking Cprior solution from ASD, decomposing with svd, and then working in U.T*sqrt(S)*U basis to pick variances
• using LU decomposition will get you better logDet
• matlab inv(Cprior) is bad --- sum(sum(inv(Cprior)*Cprior)) ~= 10e7 * nLags
  • condition number is 1e13
  • in python my sum ~= nLags + 0.07, even with that high condition number
• because of the above, bad log-evidence calculation in normal calculation (i.e. no svd no woodbury)

Mine

ro=11.03659785, ssq=1.3216941, delta=11.4020954
evidence=-10578.8521008
neg. log likelihood=57131.883003

--ro=11.0495055224 --ssq=1.32170258015 --delta=11.2919173548
evidence=-11460.235679
neg. log likelihood=3810710.97249

ro=11.0119872622, ssq=1.32220478899, delta=11.4256180793
evidence=-11500.0767877
neg. log likelihood=86630.1595886

ro=11.0680962085, ssq=1.31513284948, delta=12.8497981949
evidence=-9777.99603406
neg. log likelihood=85714.9535736