Explore reasons behind slight discrepancies between impedances from the untreated and treated networks

[cjacobscrisioni]

I have done some tests to compare OD-level impedance results from doing a Dijkstra_m on the connected, non-treated network (ie /NetworkSetup/Base_Analysis/NetwerkSpec/CreateInitialWorkingNetwork/LinkSet) and the treated, compressed network (ie /NetworkSetup/Base_Analysis/NetwerkSpec/CreateMoreEfficientNetwork/FinalLinkSet)

This is done in the /Analyses/Compression_Verification module I just pushed (db8f6e9)

Approach:

• Select a random sample of 5,000 inhabited grid cells
• Apply the dijkstra_m64 function on the above mentioned networks.
• Maximum search parameter = 360 minutes.
• Measure error as impedance with compressed network - impedance with uncompressed network

Then combine the two matrices to the cartesian product of the random sample (n x n) and export to csv file. Analyse.

Results:

• All errors are > 0, so compression leads to a structural overestimation of travel times
• Average discrepancy is 3.6 seconds on all OD links
• Errors increase somewhat with higher uncompressed travel times
• Limited outliers even up to 3 minutes

Fig above: errors on all simulated OD relations as a function of travel times with uncompressed network. Errors are compressed network travel times - uncompressed network travel times. The blue line is a floating average, see geom_smooth() in ggplot2. Black dots are individual measures.

While the discrepancies are minute, it is still worthwhile to look into them in detail to understand what the cause is.