Learning process can be distorted with the abuse of "Forget"

From SuperMemopedia
Revision as of 10:40, 27 May 2017 by SuperMemoHelp (talk | contribs) (→‎Answer: Long intervals)
(diff) ← Older revision | Latest revision (diff) | Newer revision → (diff)
Jump to navigation Jump to search

SUMMARY

Use Forget only if you change your item "beyond recognition", i.e. when the represented memory is different. Doing Forget on familiar items will result in distorting the learning process.

Problem

Daniel wrote:

I've added a difficult new item into my collection. I graded it fail many times over, until eventually I graded it pass a few times before finally grading it good. To my surprise, the item was then scheduled to appear at the same day as an easy item I've added and immediately graded good. Since this didn't make sense to me, I looked at the difficult item's repetition history and all the fail and pass grades just weren't there!

Then, I repeated the experiment with an item added and then marked as forgotten, so all those fails and passes will occur in the normal learning stage rather than the final drill stage. Now, the first grade I gave does appear in the repitition history, but not all those after that! I assume this has to do with the fact that only the first grade is given in the normal learning stage, while all the others are in the final drill stage (even if the first grade was fail...).

In both cases, the item's difficulty was not affected whatsoever. What's going on here?

Hints

  • final drill repetition is a short-term memory test. This is why it does not add to repetition history. Ideally, the user should use good item formulations to avoid final drill. For difficult material, e.g. learning Chinese, and for new users, final drill provides a degree of comfort, but does not significantly contribute to learning
  • first repetition after forgetting will often take place in a very short interval, e.g. one day. In that sense, easy and difficult items may get the same next repetition date

Follow up

Had the first repetition interval been just one day, I wouldn't have been surprised. But what I'm seeing is definitely broken behavior: when I create a new item in my collection, the first repetition interval is 33 days, and that's even if I set the FI to 3%!

Note that if I click on forget on the new item and then learn it again, it doesn't even matter whether I grade it great or bad -- the new interval is still 33 days!

Note that I've clicked on compute right before creating the new item. If I click on reset before compute, the situation is slightly better but still doesn't make any sense - the initial interval is now 11 days.

I've sent you the collection to aid in the investigation.

Answer

  • your first interval of 33 days is a result of good performance at month-long intervals. At 30 days, you still remember 94% of items. You should not play with Forget because it is equivalent of "cheating" SuperMemo. You feed the same item as it was a new item, while your performance is good. You may end up improving your memory of those "forgotten" item and make SuperMemo think you need longer and longer intervals.
  • the grade at memorizing stage does not affect the interval. it is only used to determine items that should enter the final drill

It seems the problem comes from your use of Forget. The best strategy now is to avoid using that function, and manually shorten intervals that you worry are too long. If your performance on shorter intervals is not as good as before, the interval will get shorter. It may take some time though, you have already amassed 6869 repetitions that determine the first long interval. Naturally, you could shorten that interval also by scoring poorly at first repetition, but if you do it by “cheating” you risk more problems further on. With SuperMemo, honesty is always the best policy.

I need FI=3%

I know that forgetting index of 3% is not recommended, but since my classes often had surprise quizes that I didn't know about in advance, my past experience has shown that the lowest FI is very beneficial in such cases. I would have preferred a higher value in all other cases.

Extra review

Your best warranty are good memories, not fluency. At 3% you risk crowding your learning process and slow down developing good memories. A risky, but better strategy, would be to do some random review on your memorized set, e.g. on a day before the class. This may slow down your progress in the long-term, but should ensure better recall at unexpected tests (higher fluency).

Recovering

Can I fix the collection by finding those items and resetting their repetition history, followed by resetting the SInc matrix?

Honesty is the best policy

Destroying the record of past repetitions could make things worse. SuperMemo 17 is actually less forgiving than prior versions. In SuperMemo 16 you could ditch all memory data without much effect. In SuperMemo 17, all data will be restored on first Compute in memory functions. If you delete all repetition histories, you could make SuperMemo "think" your memory is perfect. The only sensible way to recover now is to stick to the rules and overwhelm "dishonest data" with "honest data".

Optionally, you can also start a new collection to see how it performs when you avoid using low forgetting index and Forget. Once your tests are over, you could Mercy the old collection into the future and merge it with the new collection.

Followup: Too long intervals

Most of the items I currently have in my collection are items I've remembered well for quite a long time before adding them to SuperMemo. That's why I have long strings of success with them. It's too bad that they (and the use of the Forget function) have influenced SuperMemo's estimation of my memory so heavily, because my experience with truly new and difficult items is still the same as before: I might remember them correctly the next day after learning them, but after that I need to actively recall them within a few days in order not to forget them. With the intervals SuperMemo currently gives me for new items, this won't happen. Do I understand correctly that the only option I have is to leave existing items as they are, and manually shorten the intervals for all new items, repeatedly after each repetition (since currently both success and failure put the next interval way too far in the future), until they gradually return to sanity?

Answer: Long intervals

Intervals would be "too long" only if your measured forgetting index started growing above 10% (if the default requested forgetting index is lower, SuperMemo may still cause more forgetting, for many reasons). However, it is natural to have "bad feelings" about long intervals for some or even many items. Experienced users often have very good intuitions about which item is likely to depart from "the norm". This is why setting up your own intervals manually is not unusual. If it makes you feel better or enjoy learning more, you should not worry about manual intervention. The new algorithm should cope with it ok. At worst, some other important items will be crowded out from your memory. However, if SuperMemo fails to meet your measured forgetting index targets, you need to report it. You might re-check that at SuperMemopedia, but such reports probably never happen or happen rarely due to an error or a misunderstanding. In short, SuperMemo sticks to the forgetting index criteria pretty well.