SM3+ myth

From supermemo.guru
(Redirected from SM3+)
Jump to navigation Jump to search

In the context of SuperMemo, SM3+ myth is my name for the myth that Algorithm SM-5 is allegedly inferior, less stable, less predictable, or less convergent than Algorithm SM-2. A solid implementation of Algorithm SM-5 is superior in all algorithmic metrics as confirmed by practice and theory. The only conceivable weakness of Algorithm SM-5 is its higher complexity. Increased complexity could increase cost of implementation, chances for bugs, chances of low-quality implementation, etc. In that sense, all successive SuperMemo algorithms add complexity, even if the underlying models get simpler and more solid. The myth originated in Anki manual many years ago. It showed up at SuperMemo entry at Wikipedia on June 6, 2011 (good faith user Nbarth), where it resides to this day (retrieved Aug 7, 2018):

Criticism of SM3+
The SM-2 algorithm uses the performance on a card to schedule only that card, while SM-3 and newer algorithms use card performance to schedule that card and similar cards. The additional optimizations sometimes yield perverse results – answering "hard" on a card may yield an interval longer than answering "easy" on a card – and are criticized as reducing the robustness of the algorithm, making it more sensitive to variations – non-uniform difficulty of cards (a problem in versions 4 to 6, according to Wozniak), inconsistencies in studying, and so forth

I chose the name SM3+ as it seems to perfectly tag the myth. There has never been SuperMemo "Algorithm SM-3". All the said reasons for "perverse results" are false and probably originate from distorted interpretation of SuperMemo FAQs. For example, various forms of matrix smoothing are used in all algorithms, starting with Algorithm SM-5 to make sure higher difficulty converts to shorter intervals independent of the volume, and reliability of the available data. Non-uniform difficulty of cards probably means that new algorithms use absolute item difficulty, while SM-4, SM-5, and SM-6 used relative difficulty. This means that in Algorithm SM-2 items do not benefit from collected data (the algorithm does not build a model of student's memory). Items in SuperMemo 5 benefit from data biased by a given collection averages. Items in SuperMemo 8 and later, benefit from all data with no bias.

For details see this text on Superiority of SM5 over SM2.

This glossary entry is used to explain "History of spaced repetition" by Piotr Wozniak (June 2018)