Closed fasiha closed 1 year ago
tback
(t'
, t-prime) will be >= beta before updatetback
increases a lot, so you might need to go to huge tback
to get to beta = next integerClosed this because Ebisu v3 is moving away from Beta priors on recall to Gamma priors on half-life.
Interesting factoid:
for
b>=2
integer (anda
andt
arbitrary), at least forb>=2
andb<=6
.(Apologies for mixing Mathematica and Python notation.)
This might be immediately useful if
predictRecall
checks forb
(small?) integer.This can be even more useful if we can find a way to make
b
always integer in our Ebisu models—easy to do when a quiz is a success, but a bit harder to do when a quiz is a failure: basically we have to searcht'
(t-prime in the readme) that makesb
an integer.Then, predicting recall at any given time is a rational polynomial, very fast to calculate, no need for Gamma or Beta functions. We could always try to make
b=2
even.(One question I have is, right now with version 1.0, we rebalance the model to roughly near the half-life so
a
andb
aren't too different. Are theret'
s that yield Beta distributions that are more faithful to the GB1 posterior than others?)