Tweak SlimStampen parameters
dljve opened this issue · comments
Currently the tiles decay too fast.
- Are the current model parameters (n, c, f, F, t) adequate?
- Is the standard alpha of 0.3 adequate?
- In the decay calculation in
updateModel: (1) how to handle the rare case of division by zero (2) why is the negative log sometimes taken in the nominator? Is this mathematically plausible (see Nijboer 2011, page 13) or is there an error in the implementation?
- n=15, c=0.25, f=0.3 and t=0.5 are standard values used in the papers. We changed F from 1 to 0.8 but it shouldn't make much difference.
- The fast decay only happens at the start. A probable solution is to slightly lower the initial alpha value, e.g. a=0.29
- This will only affect optimal alpha estimation, it shouldn't cause the issue?
The issue was the initial alpha. Now set to a = 0.25
This was probably caused by a bug regarding the intertwined use of both server time and client time.
See commit 5eee4152e3f91bc087d76a50e3f3c3a21a76aac3. Reverted the initial alpha to 0.3.