I’ve enjoyed playing with the FSRS simulator in the latest Ankidroid build. One interesting thing I’ve noticed is that with a fixed review count per day, lower target retention tends to lead to higher learning rate even when you adjust the total reviews per day to ensure total time doesn’t increase with the lower retention rate setting. I found the learning rate seems to keep increasing, suggesting an optimal retention rate below the minimum of 70% the current simulator allows.
Here are two simulations, the first is with 90% target retention and 70 reviews per day, the second is 70% target retention and 50 reviews per day, which the simulator believes leads to approximately the same time spent.
Is this expected? Or is this an indicator that the simulator is not correctly estimating the time of failed reviews with the lower target retention rate?
1 Like
That’s intended, so we remove CMRR in the latest release.
1 Like
<70% optimal retention rate differs significantly from what CMRR recommended me before (iirc around 85-90%), even when I’m reducing the reviews/day with the lower target retention rate to make total time approximately match (which iirc CMRR doesn’t do).
Am I also right to assume the simulator is correctly simulating learning/releasing steps after failed reviews?
This seems a bit odd to me. Also, given the simulator is suggesting a rate below even 70% for me, does it make sense to decrease the minimum target retention the simulator allows?
1 Like
Because FSRS-6 adopts a flexible forgetting curve, the optimal retention is usually lower than 70%.
2 Likes
Am I also right to assume the simulator is correctly simulating learning/releasing steps after failed reviews?
Yep
This seems a bit odd to me. Also, given the simulator is suggesting a rate below even 70% for me, does it make sense to decrease the minimum target retention the simulator allows?
That would also involve reducing the limit for the actual desired retention.
The current plan with CMRR is to replace it with a workload graph, so that will hopefuly make it obvious where the optimal workload is.
2 Likes
