Once fixing #241 , and bumping CONF_MULTIPLIER to 3 we experience the following things that need to be fixed:
- The classification multipliers are more modest. This can be fixed by copy-pasting the dataset (e.g., 10 times)
- The Forest CovType results for the MAB solver are still worse than for the
EXACT solver. This is consistent with what we wrote in the paper, but we should investigate
- Get tests to pass
- Some of the logarithmic scaling plots look worse on MNIST. This is probably due to the data deduplication issue (fixing 4x MNIST -> 1x MNIST)
Once fixing #241 , and bumping
CONF_MULTIPLIERto3we experience the following things that need to be fixed:EXACTsolver. This is consistent with what we wrote in the paper, but we should investigate