Ecosyste.ms: Awesome

An open API service indexing awesome lists of open source software.

Awesome Lists | Featured Topics | Projects

https://github.com/davidgilbertson/disco-mundus

Learn the world - map-based memory training
https://github.com/davidgilbertson/disco-mundus

Last synced: 13 days ago
JSON representation

Learn the world - map-based memory training

Awesome Lists containing this project

README 

        
A memory-training app, on a map.



https://discomundus.web.app



Currently, this only has data for Sydney suburbs. Coming soon (maybe): countries

of the world



PRs are welcome if you have the GeoJSON. Open an issue to discuss.



# Spaced repetition algorithm



I considered copying something like

[SM-2](http://super-memory.com/english/ol/sm2.htm), but did not.



My implementation is (I think) better suited to these strictly-geographical

questions where distance plays an important role. Also, there is less variation

in ease-of-recall between different questions, compared to the questions in a

general purpose memory app, so the 'Easiness Factor' that SM-2 uses would be of

less benefit.



## The Disco Mundus algorithm



### Short version



It doubles the interval for correct answers and allows for partial correctness.



### Long version



First, a score between 0 and 100% is generated for the user's answer.



- If they were correct, let the score be 100%; else

- If the feature they tapped is adjacent to the correct feature, let the score

  be 80%; else

- The distance between the tapped place and the center of the correct feature is

  used to grade the score:

  - Between 0 and 4KM, grade the score from 100 down to 0%.

  - x KM or more away, let the score be 0%



The score is then converted to a multiplier between 0 and 2.



The time since the last review (or 10 minutes, if this is the first review) is

then multiplied by the multiplier, and used to set the next review time.



There is a minimum review interval of 1 minute.



Answering incorrectly essentially 'resets' a question, since the multiplier is

0, it will revert back to 1 minute.



### Session queue



At all times, questions are asked from a queue. This is a subset of the total

set of questions. All questions in the session queue are presented for review

before any new questions are asked. When the page is first loaded, any questions

ready for review are put in the session queue. Once all of these questions have

been reviewed sufficiently (and the session queue is empty) 10 new questions

will be picked to populate the session queue (to be configurable in the future).



A question may be presented several times, based on the next review date and the

lookahead window. For example: a user is reviewing their questions, and answer

one incorrectly. The 'next review time' is set to one minute. Because one minute

is less than the 'lookahead window' of 10 minutes, it is kept in the queue so

that it can be reviewed again before moving on to new questions. Specifically, a

question answered incorrectly once must be answered correctly twice before it's

removed from the session queue.



If a question is answered correctly on the first attempt, its next ask date will

be 20 minutes in the future, which falls outside the lookahead window, and so it

will be removed from the session queue immediately, avoiding unnecessary

repetitions.



### Variables



- **Multiplier:** for a correct answer, how much will the previous interval be

  multiplied by to get the next interval.

- **First review time:** when a question is first answered, what should the

  review time be for a correct answer.

- **Minimum review time:** If a question is answered incorrectly (has a score of

  0%), the next review should not be set to 0 minutes. Instead, this minimum

  should be enforced.

- **Lookahead window:** when selecting the next question, how far into the

  future should we look. E.g. if lookahead window is 10 minutes, and there is a

  question due for review in 5 minutes, then the system would ask that question,

  before removing it from the queue and moving on to new questions.



### A weak spot



A user may know and answer very well, and the last duration was a few months,

but they bump the screen and get the answer wrong. The question gets reset, but

really it will only be five or six extra reviews to get it back up again.



## Ongoing review



The results will be monitored. The goal is that 90% of answers are correct or

very close (score > 80%) when using the app daily - dependent on how many new

questions are being asked.



For the well-remembered questions, there would be ~4 revisions in the first

week, 2 more for the rest of the month, another 4 for the rest of the year, and

so on.



Running `printStats()` in the console will print stats. Also clicking on my

house will show them as a popup :)



The scaling of the score is currently hardcoded as 4KM, as the app only shows

Sydney suburbs and this scale works fine. Eventually, I'll want something more

sophisticated like the distance between two points as a ratio of the diameter of

the target feature.