order-dispatching-fleet-management
This is the final project of my algorithm course. In this project we use a min-distance greedy + maxflow-based fleet management algorithm to improve the accumumalted income(ADI) and order repronse rate(ORR).
Requirement
- python 3.6
- numpy
Explain for each file
- /pic: the result diagrams
- /report: our report
- config.py: the config file. you can change settings like dirver number here
- data_process.py: autoprocess data for simulation
- gps_utils.py: change gcj postion into wgs84 positiion
- main.py: the main function is here. You can change policy method in main
- maxflow_solver.py: a class to solve maxflow problem. But in the project the maxflow problem is special, I didn't implement this solver in standard way.(Which means you can't use this to solve other problem). Good news is this is much faster.
- policy.py: Here includes all policies:random and greedy. You can add new policy success from the base class.
- simulator.py: a simulator to simulate the real world
- utils.py: some utils function
- view.py: visualize the results and save pic to /pic
- README.md: which you are looking :)
Clean Data
Data can be download from didi
You just need to put data in the right position and run
formalize_order_data()
sort_orders()remove wrong time
remove far position
total removed: 131566
remaining orders: 6934371
Grid
about 2km x 2km for each grid, 15 * 17 grids
range: [[103.89325, 30.547255], [104.20675, 30.852745]]
split by degree, x(0.0209), y(0.01797)
Time
time step is 10 mins(600s), all order which are finished in less 1 time step are setted as 1 time step. (About 230281 orders)
Order
price
start step : 8
base dis: 2 km
price: 1.9 yuan/km
night: 23:00 - 6:00 +20%
delay
each order lasts 1 time step
Driver Bias
About 4,000 drivers in one day. But they won't be online all day. So we add a bias to driver number.
n_driver = bias * n_driver
unman car = n_driver * ratio_unman(ratio)
Results
ADI, ratio_unman = 0.5
| bias | 0.1 | 0.3 | 0.5 | 0.7 | 0.9 |
|---|---|---|---|---|---|
| random | 199.99 | 306.92 | 337.27 | 354.23 | 366.05 |
| greedy | 205.76 | 294.85 | 326.90 | 346.86 | 359.26 |
| greedy+fm | 268.15 | 382.73 | 388.18 | 391.13 | 393.99 |
ORR, ratio_unman = 0.5
| bias | 0.1 | 0.3 | 0.5 | 0.7 | 0.9 |
|---|---|---|---|---|---|
| random | 0.4843 | 0.7484 | 0.8260 | 0.8689 | 0.8996 |
| greedy | 0.6357 | 0.8024 | 0.8613 | 0.8973 | 0.9196 |
| greedy+fm | 0.7683 | 0.9692 | 0.9772 | 0.9815 | 0.9857 |
In greedy there are more idle time for each car So the ADI is lower than random.
bias = 0.3, ratio_unman = 1, greedy+fm
| driver ratio | 0.1 | 0.2 | 0.3 | 0.4 | 0.5 | 0.6 | 0.7 | 0.8 | 0.9 | 1 |
|---|---|---|---|---|---|---|---|---|---|---|
| ADI | 103.83 | 192.17 | 269.16 | 328.31 | 361.48 | 373.56 | 378.58 | 381.47 | 383.26 | 384.05 |
| ORR | 0.3970 | 0.6205 | 0.7737 | 0.8797 | 0.9359 | 0.9554 | 0.9633 | 0.9675 | 0.9701 | 0.9712 |
It show if we replace all cars with unman cars, keeping about 86% car numbers can get the same performance.
And even you use only 50% cars the performance is still good.(Unman is the future!)
To do
fleet management