First column = tune iteration, second column = tune BLEU

Choosing the average of a few random sentence's model->hope direction, With Vlad's default --bleu_weight=1, nothing improved, and convergenced seemed to be slower, which is the opposite of what we intended.  5 random directions and the orthogonal directions were still used (fewer than normal random directions; the oracle directions replaced some of them).

dev.bleu.baseline.txt - usual search directions (orth+random)

dev.bleu.oracle_mira_direction.txt - as above, but fewer random directions, 10 model->hope directions

note that the search directions are recomputed after each MERT iteration.

/home/jgraehl/tune/urdu/vest-dth-oracle*

~/tune/urdu$ ls -d vest-dth*
vest-dth  vest-dth-oracle  vest-dth-oracle-bleuwt=0.1  vest-dth-oracle-bleuwt=10  vest-dth-oracle-bleuwt=100

oraclewt devtest-estimated-BLEU
0   22.01 (orth. dirs, 10 random dirs)
0.1 22.13 (orth. dirs, 10+10 model->hope and fear->hope random directions, 5 purely random dirs)
1   22.06 "
10  21.85 "
100 22.08 "

no evidence of any advantage using oracle directions (TODO: decode mt09 as test set, check for any advantage when using many features)