driving time impedence?

GeoDaCenter / spatial_access

https://spatial.uchicago.edu

37 stars 11 forks source link

driving time impedence? #45

Closed ifarah closed 5 years ago

ifarah commented 5 years ago

Comparing Dan's matrix, a difference of 10 minutes on average.

jkoschinsky commented 5 years ago

our driving times are 10 minutes lower than OTPs and were below the lowest time threshold in Google for the test cases we looked at (for non-rush-hour times): can we up our impedance?

lmnoel commented 5 years ago

Interesting observation: There are many segments in the driving network that are just a few meters long. Because I'm using integers for these values (as opposed to floats, which OTP uses I think), when divided by high speeds, these edges turn in to 0. I'm still trying different combinations of node impedance/edge rate to account for this best--but this seems to be the root of the difference.

ifarah commented 5 years ago

@lmnoel , did you get it working yesterday or is the problem elsewhere?

lmnoel commented 5 years ago

Didn't have enough time yesterday. I'll be back in this afternoon.

ifarah commented 5 years ago

awesome, thanks! let me know if I can help with anything

lmnoel commented 5 years ago

Actually @ifarah, I think I have a slightly different version of the tracts dataset than your tracts2010.csv file. Could you send me that exact one?

lmnoel commented 5 years ago

Using the tracts Irene sampled, the original edge weight model of p2p gave the following (first value is mean(dan's time - p2p's time) in seconds, second value is std dev:

diff_17031836200
972.1290726817043
369.5565549537744
diff_17031071100
779.3771929824561
346.8469111225043
diff_17031561100
1137.9649122807018
348.7904301116148

When I step down each edge weight to 40% of its original value and add a 4s node penalty, we get the following:

diff_17031836200
149.15288220551378
91.72946796319982
diff_17031071100
5.053884711779449
113.99102235752898
diff_17031561100
33.177944862155385
127.18073532449111

original model original model diff original_diff modified model modified model diff modified_diff

The correlation between dan's predicted time and error was crazy high in the original model (r>0.99). In the modified version it drops to r=0.58.

So the quality of the model seems to be entirely dependent on the quality of the actual driving speed data that is given as input.

lmnoel commented 5 years ago

I'll push up these changes in the next hour or so

ifarah commented 5 years ago

It totally makes sense that the quality of the model depends on quality of driving speed right? That would be in the config settings? or where are you tweaking that?

lmnoel commented 5 years ago

yeah

ifarah commented 5 years ago

Are you still working on this, @lmnoel ?

lmnoel commented 5 years ago

Not currently, and I have no plans to unless you need anything further