For branch https://github.com/SANDAG/ABM/tree/ABM3_bike_route_choice_RSG
Logsums in the new bike route choice model are consistently lower (higer cost) than logsums in the Java bike model. Looking at MGRA, there is a mean logsum difference of -2.6, with standard deviation 1.3:
Here is a random sample of 20 MGRA pairs, all with lower logsum in Python. Note that a few of these had positive logsums in Java, but there are no positive logsums in Python:
Here is an example MGRA pair (24297, 23637) in the Python model:
Line color is iteration, line width is cost. The paths generated seem reasonable (the high-cost pink line is a very steep climb, 8 of 10 paths avoid it), and the bike time is very close between Java and Python. The major difference is the final logsum, which implies to me a difference in the utility calculation. Being able to trace the same MGRA pair in Java would be helpful here for a direct comparison.
For branch https://github.com/SANDAG/ABM/tree/ABM3_bike_route_choice_RSG
Logsums in the new bike route choice model are consistently lower (higer cost) than logsums in the Java bike model. Looking at MGRA, there is a mean logsum difference of -2.6, with standard deviation 1.3:
Here is a random sample of 20 MGRA pairs, all with lower logsum in Python. Note that a few of these had positive logsums in Java, but there are no positive logsums in Python:
Here is an example MGRA pair (24297, 23637) in the Python model:
Line color is iteration, line width is cost. The paths generated seem reasonable (the high-cost pink line is a very steep climb, 8 of 10 paths avoid it), and the bike time is very close between Java and Python. The major difference is the final logsum, which implies to me a difference in the utility calculation. Being able to trace the same MGRA pair in Java would be helpful here for a direct comparison.