Features in next version of Parley
inge at lysator.liu.se
Tue Aug 26 20:56:07 UTC 2014
I did know about the forgetting curve before but I haven't used it as explicitly as you did
here. I think the leitner method is trying to do the same as you are but using much less
sophisticated and detailed tools. Remember that Leitner came up with his system when
computers didn't exist and people were forced to use physical cards with words printed on
them. We can indeed do better now.
I am not sure if you remember when I was discussing whether we should implement time
based or leitner box based scheduling of words. The time based approach I was thinking of
tries to do about the same that you are describing here but without estimating the
"forgetting constant" for each individual. Instead I went with the time intervals that were
already in parley which I suppose is some form of guess/approximation of what the curve
looks like for the average person.
There are some things I don't understand in your description below. For instance how you
calculate the optimal intervals for an individual? Or do you? How is the algorithm affected
if you don't train for some days even though it is scheduled? How is it handled when you
introduce new words in a collection that is already trained some time? Etc.
But in general I think if this works as well as you say, it would be a wonderful
enhancement! Let's discuss this more.
As a side note, I don't agree that dropping to 0 is discouraging - but then I have now ego
connected to the bars being fully green. :) I can totally understand how some people
would be very irritated by having a fully green bar and then one or two words in the yellow.
I only see the it as a sign that there are words that I nee to train again, nothing more.
But as you also note, it's a fine line between optimal learning and being discouraging. I
think we want to be on the safe side and encourage the student to train more rather than
be optimal at the cost of disillusionment.
On Tuesday, August 26, 2014 13:27:20 Andreas Xavier wrote:
> Feature: Forgetting curve based training schedules
> Note: The version with the nice .svg plots got bumped.
> I want to add training schedules and user assessments based
> directly on the forgetting curve time constants.
> Below I will explain what the forgetting curve is,
> how Parley currently schedules training intervals and what the
> new method will be and why it is better.
> Please skip the first two sections if have already know how
> Parley works.
> I am not really proposing this for release until after the more
> detailed training data is in the data file.
> Forgetting Curve:
> The forgetting curve measures how fast you forget any new
> information that you learn. It looks like a slide at children's park.
> When you first learn something you very quickly forget,
> the time constant is short or fast, the slide is steep.
> Every time that you learn the material again the time constant
> gets longer, the slide gets flatter and it takes longer for you
> to forget the information.
> For more detail see https://en.wikipedia.org/wiki/Forgetting_curve
> If you want to learn new material as quickly as possible, then there
> is an optimal place on the curve/slide to relearn the material. Too
> late and there is risk that you have forgotten too much, too early
> and the gain in time constant is smaller than it could be. Too early
> is not that bad it just means that overall the learning will take longer.
> However, if you need to learn something in one week, spending 40
> hours in that one week is better than spending 10 hours over 4 months
> , even though 10 hours is less total time.
> Leitner Boxes:
> Currently, Parley uses the Leitner box method to schedule training
> intervals. The Leitner box method is an elegant algorithm to approximate
> optimal training intervals so that the student doesn't have to calculate
> time constants by hand. The method is simple, every time you get a correct
> answer you double the training interval and with every wrong answer you
> either halve the interval or go right back to the shortest interval.
> The Leitner method has no provision for intense focused study, like the
> 40 hours in one week example above. If I force Parley to allow me to
> train a word 14 times in a row, it will retire it immediately and it
> The Leitner box method unduly penalizes late words. If training of a word
> is very late, then it tells Parley very little about what the student really
> knows. But the Leitner box method penalizes the student anyway.
> I think the Leitner method as implemented is too visible in a discouraging
> way. I get one wrong answer and I drop right down to box 0. In reality
> a wrong answer teaches the student almost as much as a correct answer,
> but it indicates the student needs to review it sooner.
> At the maximum rate of learning before a student approaches mastery there
> are many wrong answers.
> New Method:
> The new method is simply to use the new training data to directly calculate
> the forgetting time constant for each word in reading, listening, speaking
> and writing.
> It is better because it can provide optimal training intervals directly even
> when the student practices early, late or often. It correctly handles
> short training intervals.
> It is easier to set up. In fact if you do nothing it will end up at optimal
> training intervals for that student.
> It can allow the student to do planning. For example the student can ask
> the question: I would like to be at 90% retention of these words 4 months
> from now and be able to remember them for 1 month. How much time should I
> study per day?
> This method and the data stored in the new file format makes available a new
> array of statistics and plots to track your performance. Unfortunately, if
> the visualizers highlight new words students will have unscheduled training
> opportunities. =)
> The method also allows interactive tailoring of the rate of learning to use
> preferences. Simply asking the user if the previous session seemed too
> hard, too easy or just right would allow the algorithm to tailor the rate
> of learning and the intervals to the user's desires. Themaximal rate of
> learning probably involves more wrong answers that most students are
> comfortable with.
> I made a demo of the method with a mock student. The code for the new
> method is real. The student is simulated.
> I have attached two plots of results of fixed interval training, where the
> student practices only the words due up to some time limit each session.
> The first plot is with 10 words over 1 month, 9 training sessions of 30
> seconds each. You can see the improvement of the time constant by the
> flattening of the curve at the end of the month.
> The second plot is of 1000 words, 1/2 hour per day over 6 months.
> You can see the gain and then maxing out of performance. The curve flattens
> out before 90% because I stuck in a fudge factor so that I could complete
> the example before the end of today.
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the kde-edu