Lower-bound RAPTOR

[mority]

• creates lower bound adjacencies per location during timetable import
  • locations that are reachable without transfer and the respective minimal travel time
• before doing the actual routing, run some sort of RAPTOR on the lb_adjacency graph to determine the minimal traveltime per number of transfers for each location from the destination

Overhead of LB adjacencies:

• Calculating them during import takes 2.3 s on the DELFI GTFS
• the file size of the serialized timetable increases by ~30%

Running time [ms] of LB RAPTOR for 1000 random Queries on the DELFI GTFS:

count  996.000000
mean   360.838353
std     54.876334
min    238.000000
25%    322.000000
50%    358.000000
75%    397.000000
max    615.000000

(I excluded 4 values = 7 ms. Probably, involved some island destination not connected to the rest of the timetable)

The main driver seems to be slow convergence. Regardless of the destination LB RAPTOR performs the maximum number of rounds. Even k = 15 yields minor improvements in travel time and therefore station marks are set until then. The current version takes transfer time into account and adds it to the travel time which helps a bit with convergence since it adds up and makes it harder for later rounds to trigger station marks with one-minute improvements or such.

Evaluation on DELFI GTFS

Responses equal: 959/1000

Baseline

response_time
average: 394.78
max: 4,868
99 quantile: 2,549
90 quantile: 1,080
80 quantile: 615
50 quantile: 155
min: 0

LB_RAPTOR

response_time
average: 1,393.15
max: 6,941
99 quantile: 4,567
90 quantile: 2,291
80 quantile: 1,708
50 quantile: 1,147
min: 0