Open fsibapt opened 1 month ago
In the documentation [1] you will see the difference between the filtering procedure (using the last posterior as prior of the next event), and the smoothing procedure (propagating the information thought the entire Bayesian network). Filtering is "online", but in a sub-optimal way, because we are not using all the historical information. The best estimates are achieved with the smoothing procedure, but requires an iteration to reach convergence, so in general it is used only "offline". If the online
parameter is set to true
, then the smoothing procedure will be executed at each time step. It is inefficient but uses the best posterior estimate as prior of the next event at each time step. The main reason for adding this parameter was experimental, to evaluate the (best) prior predictions made by the model.
[1] https://github.com/glandfried/TrueSkillThroughTime/releases/download/doc/landfried-learning.pdf
I've noticed that in the julia implementation the History class includes a paramater called "online" that does not exist in the R or python versions, I also couldn't find anything about it in the documentation. What is its purpose and how is it used?