Mobility-on-demand fleet management application

Installation

Locate the transportation environment module at file tenv\network.py:

tenv_mod = "C:\\Users\\LocalAdmin\\OneDrive\\leap_forward\\street_network_server\\tenv"

ADP config

state = TIME, LOCATION, BATTERY, CONTRACT, CARTYPE, CARORIGIN decision = ACTION, POSITION, BATTERY, CONTRACT_DURATION, CAR_TYPE, CAR_ORIGIN, ORIGIN, DESTINATION, SQ_CLASS, N_DECISIONS

Loading intances from files

Indicate the instance file by setting variable:

instance_name = f"INSTANCE_PATH/exp_settings.json"

Running the simulator

The command below will start the simulation on the port 5002.

bokeh serve --show --port 5002 tests\ml\live_simulation.py

The time between each assignment can be set by changing the variable STEP_DELAY (seconds) in the PlotTrack class.

Running the service rate and fleet status graphs

bokeh serve --show --port 5003 mod\visual\slide_episode.py

Creating the environoment

Execute the following commands:

conda config --prepend channels conda-forge
conda create -n env_slevels python=3.7
conda activate env_slevels
conda install --strict-channel-priority osmnx
conda install gurobi
conda install gurobipy
conda install -c conda-forge --strict-channel-priority h3-py
conda install bokeh
pip install seaborn

To activate the environment use:

C:\Users\LocalAdmin\Anaconda353\Scripts\activate env_slevels
C:\Users\breno\Anaconda3\Scripts\activate env_slevels

Execution

Logging

Keyword	Function
level LOG_LEVEL	Choose logging level LOG_LEVEL = [INFO, DEBUG].
log_mip	Save mip model (`.lp`) and mip solution log (`.log`) for each time step and iteration.
log_summary	Show fleet stats for every step.
log_all	Log all steps (MIP decisions, dual extraction, and VFA update).
log_fleet	For each iteration `n`, save `cars_n.csv` and `cars_result_n.csv` log files in `fleet/fleet_data/` comprising the information of each car in the beginning and in the end of iteration `n`, respectively. The following fields are considered for each vehicle: `id, type, point, waypoint, previous, origin, middle_point, elapsed_distance, time_o_m, distance_o_m, elapsed, remaining_distance, step_m, idle_step_count, interrupted_rebalance_count, tabu, battery_level, trip, point_list, arrival_time, previous_arrival, previous_step, step, revenue, n_trips, distance_traveled, status, curret_trip, time_status, contract_duration`.
log_trips	For each iteration `n`, save (1) `trips_n.csv` and (2) `trips_result_n.csv` log files in `trip_samples_data/` comprising the information of each trip request in the beginning and in the end of iteration `n`, respectively. The following fields are considered in (1): `placement_datetime, pk_id, dp_id, sq_class, max_delay, tolerance, passenger_count, pickup_latitude, pickup_longitude, dropoff_latitude, dropoff_longitude`. The following fields are considered in (2): `placement_datetime,pk_id,dp_id,sq_class,max_delay,tolerance,passenger_count,pickup_latitude,pickup_longitude,dropoff_latitude,dropoff_longitude,pickup_step,pickup_delay,pickup_datetime,dropoff_time,dropoff_datetime,picked_by`.
log_times	Log the times (per iteration) to create decisions, extract duals, realize decisions, update duals, setup costs, setup constraints, and optimize MIP. Also include the total time.
save_plots	Save iteration's demand and fleet statuses for each step.
save_df	Save .csv dataframes with fleet (status, vehicle count) and demand (met, unmet) data

Scenario configuration

Keyword	Function
hire	Consider FAV hiring. Related settings: DEPOT_SHARE, FAV_FLEET_SIZE, MAX_CONTRACT_DURATION.
n "number of iterations"	Set the number of iterations.
FLEET_SIZE "number of vehicles"	Set the fleet size.
backlog "extra_waiting"	Activate user backlogging. Rejected users are re-inserted into the demand for `times = extra_waiting/time_increment`. E.g., if the period is `1 min` and a unmatched user can wait up to `10 min`, this user is re-inserted in the demand pool throughout the next 9 iterations.

Optimization methods

The folowing keywords can be used to define the optimization method to assign and rebalance vehicles. For a single instance, each keyword creates a separate folder where you can find the results of each method.

Keyword	Method
train "n"	Run ADP algorithm on the training dataset to create VFAs. Update the file progress.npy with the updated results for each visited state every "n" iterations (by default, "n" is 1).
test	Run ADP algorithm on the testing dataset and uses the VFAs (from training) to measure the impact of future decisions.
myopic	Barebone assignment algorith with no rebalancing.
policy_random	Assignment algorithm with random rebalancing.
policy_reactive	Assignment algorithm and Alonso-Mora rebalancing (send vehicles to unmet requests but interrupt rebalancing at each period to check whether new requests can be picked up).

Demand

Enable USE_CLASS_PROB to load probabilities of picking up users of different classes throughout time and space. If disabled, user membership to classes is defined at random according to the proportions defined in TRIP_CLASS_PROPORTION. For example, to consider 80% of the demand is comprised of "B" users and the remaining is comprised of "A" users, we can do as follows:

TRIP_CLASS_PROPORTION: (("A", 0), ("B", 1))

Execution

Configure the json and run it passing folder and file. For example:

main.py 'c:/tud/mod/config/adp_tune/' 'TS_2K-B-1-1.json'

Batch file (.bat):

call C:\Users\LocalAdmin\Anaconda3\Scripts\activate env_slevels
call cd C:\Users\LocalAdmin\OneDrive\leap_forward\phd_project\reb\code\mod\
call python mod\ml\adp_network_server.py R -FLEET_SIZE 300 -n 500 -train -save_plots

Memory profiling

Decorate the function you would like to profile with @profile and execute passing the option -m memory_profiler (e..g, python -m memory_profiler tenv\util.py)

Printing functools

print(amod.cost_func.cache_info())

Installing cython

https://wiki.python.org/moin/WindowsCompilers

Execution: python setup.py build_ext --inplace

MIP

Gurobi error codes

Once an optimize call has returned, the Gurobi optimizer sets the Status attribute of the model to one of several possible values. E.g.:

Status code	Value	Description
LOADED	1	Model is loaded but no solution information is available.
OPTIMAL	2	Model was solved to optimality (subject to tolerances) and an optimal solution is available.
INFEASIBLE	3	Model was proven to be infeasible.
INF_OR_UNBD	4	Model was proven to be either infeasible or unbounded.

Gurobi method

Algorithm used to solve continuous models or the root node of a MIP model (info). Options are:

-1=automatic
0=primal simplex
1=dual simplex
2=barrier
3=concurrent
4=deterministic concurrent
5=deterministic concurrent simplex

Example:

m.setParam("Method", 1)

Gurobi MIP focus

If you believe the solver is having no trouble finding good quality solutions, and wish to focus more attention on proving optimality, select MIPFocus=2. If the best objective bound is moving very slowly (or not at all), you may want to try MIPFocus=3 to focus on the bound (info).

If you are more interested in finding feasible solutions quickly, you can select MIPFocus=1.

Example:

m.setParam("MIPFocus", 1)

AMoD

Online and loaded data

If you define amod = AmodNetworkHired(config, online=True) the following is activated:

# Is data calculated on-the-fly?
# Calculate everything on the fly
if online:
    self.revenue = self.online_revenue 
    self.cost = self.online_costs
    self.penalty = self.online_penalty

# Access data from dictionary
else:
    self.load_od_data()
    self.revenue = self.loaded_revenue
    self.cost = self.loaded_costs
    self.penalty = self.loaded_penalty

If online=False, revenue, cost, and penalty data are loaded into dictionaries.

Logged information

Keyword	Description
step=NUMBER	Indicate the start of step.
Reachable	List of reachable trips.
Unreachable	List of trips that cannot be reached by any vehicle.
TRIP COUNT ORIGIN	Number of trips per origin location.
TRIP COUNT DESTINATON	Number of trips per destination location.
LOG COSTS [DECISION, SOLUTIONS]	Cost calculation of decision set(DECISION -> COST + DISCOUNT * POST = TOTAL).
SOLUTIONS	DECISION = TIMES APPLIED
Car attributes	List of cars and attributes (id, position)
NODE - NEIGHBORHOOD	Table: node_id, levels_ids, unreachable, neighborhood
ATTRIBUTE CAR COUNT	Table: g, step, inbound, cars, unrestricted

Methods

Reactive rebalancing

1 - Match idle/rebalancing cars to trips (ignore rebalancing decisions)
2 - Update fleet, such that non-matched cars that were previosly rebalancing, continue to do so.
3 - Match rejected & outstanding trip origins (the rebalancing targets) to idle vehicles.

Tuning

Keyword	Method
TEST_LABEL	Label of all tuning instances (default "TUNE") .
N_PROCESSES	Number of tuning istances executed in parallel (by default, 2).
FOCUS	Key to select pre-configured tuning instances.
METHOD	Which optimization method is used in the tuning (-train, -test, -reactive, -myopic).

Example:

python tests\ml\tuning.py <TEST_LABEL> <N_PROCESSES>  <METHOD> <FOCUS>
python tests\ml\tuning.py SH 9 -train hiring

Notebooks

To compare the weight put in each level in the hierarchical aggregation process, use the file notebooks/

Data

Map

The street network consists of a strongly connected directed graph G=(N,E) comprised of N=6,430 nodes and E=11,581 edges.

Distance matrix

The distance matrix in meters (float):

dist_matrix_m.csv

The distance matrix in seconds considering a 20km/h speed (integer):

dist_matrix_duration_s.csv

Node list

The node list file nodeset_gps.json contains the following information:

{
    "nodes": [
        {
            "id": 0,
            "x": -73.9600434,
            "y": 40.7980486
        },
        ...
        {
            "id": 6429,
            "x": -73.93256054530173,
            "y": 40.860556269812896
        }
    ]
}

Node list and regional centers

The node set file nodeset_info.json contains the following information:

{
    "nodes": [
        {
            "id": 0,
            "step_center": {
                "60": 1326,
                "300": 2679,
                "600": 6392
            },
            "x": -73.9600434,
            "y": 40.7980486
        },
    ...
        {
            "id": 6429,
            "step_center": {
                "60": 1258,
                "300": 4506,
                "600": 6429
            },
            "x": -73.93256054530173,
            "y": 40.860556269812896
        }
}

The step_center attribute stores the regional center of each node, reachable within 60, 300, and 600 seconds.

First class distrubution

In order to assign service quality classes to the requests, we assume that the first-class user locations and request times coincide with the 20% most generous tippers (among tipping users) of the Manhattan demand occurring between 5AM and 9AM. To make our sample more representative, we first aggregate all demand data from 2011 Tuesdays. Then, we assign first-class labels to all requests whose tip/fare ratio ranks over the 80^th percentile (which is around 0.26).

We aggregate probabilities in 5 min bins, such that 24h = 288 bins. Example:

{ 0 [0:00 - 0:05), 1 [0:05 - 0:10), ..., 59 [5:00 - 5:05), 60 [5:05 - 5:10), ..., 107 [8:55 - 9:00), ..., 287 [23:55 - 00:00) }

The 1^st class probability file is 1st_class_prob_info.npy. The structure is as follows:

{
    "time_bin": 5,
    "start": "5:00:00",
    "end": "9:00:00",
    "data": {
        0: {
            67: 0.5,
            74: 0.5,
            82: 0.25,
            83: 0.2,
            88: 0.16,
            89: 0.18,
            90: 0.25,
            91: 0.22, 
            92: 0.2,
            93: 0.16,
            94: 0.2,
            95: 0.16,
            96: 0.16,
            97: 0.11,
            98: 0.08,
            100: 0.12,
            105: 0.22,
            106: 0.2,
            107: 0.2
        }
        ...
    }
}

The attribute data stores for each node id the probabilities associated with the appearance of 1^st class probabilities at each time bin. When a pair (node id, time bin) does not exist, the probability is zero.

Pickup and delivery requests

Excerpts from the Manhatan taxicab dataset where coordinates are approximated to the closest node in N within a 50-meter range.

Training data

The ADP estimates value functions by randomly sampling 10% of the request demand of a single day (Tuesday, 2011-04-12) in in the interval [5AM, 9AM).

The training data correspond to file:

tripdata_ids_2011-04-12_000000_2011-04-12_235959.csv

Testing data

The quality of the value functions estimated using the training data is tested against the remaining 51 Tuesdays of the year. Again, a 10% sample in the interval [5AM, 9AM) is used for comparison.

The testing data correspond to files:

tripdata_ids_2011-01-04_000000_2011-01-04_235959.csv
tripdata_ids_2011-01-11_000000_2011-01-11_235959.csv
tripdata_ids_2011-01-18_000000_2011-01-18_235959.csv
tripdata_ids_2011-01-25_000000_2011-01-25_235959.csv
tripdata_ids_2011-02-01_000000_2011-02-01_235959.csv
tripdata_ids_2011-02-08_000000_2011-02-08_235959.csv
tripdata_ids_2011-02-15_000000_2011-02-15_235959.csv
tripdata_ids_2011-02-22_000000_2011-02-22_235959.csv
tripdata_ids_2011-03-01_000000_2011-03-01_235959.csv
tripdata_ids_2011-03-08_000000_2011-03-08_235959.csv
tripdata_ids_2011-03-15_000000_2011-03-15_235959.csv
tripdata_ids_2011-03-22_000000_2011-03-22_235959.csv
tripdata_ids_2011-03-29_000000_2011-03-29_235959.csv
tripdata_ids_2011-04-05_000000_2011-04-05_235959.csv
tripdata_ids_2011-04-19_000000_2011-04-19_235959.csv
tripdata_ids_2011-04-26_000000_2011-04-26_235959.csv
tripdata_ids_2011-05-03_000000_2011-05-03_235959.csv
tripdata_ids_2011-05-10_000000_2011-05-10_235959.csv
tripdata_ids_2011-05-17_000000_2011-05-17_235959.csv
tripdata_ids_2011-05-24_000000_2011-05-24_235959.csv
tripdata_ids_2011-05-31_000000_2011-05-31_235959.csv
tripdata_ids_2011-06-07_000000_2011-06-07_235959.csv
tripdata_ids_2011-06-14_000000_2011-06-14_235959.csv
tripdata_ids_2011-06-21_000000_2011-06-21_235959.csv
tripdata_ids_2011-06-28_000000_2011-06-28_235959.csv
tripdata_ids_2011-07-05_000000_2011-07-05_235959.csv
tripdata_ids_2011-07-12_000000_2011-07-12_235959.csv
tripdata_ids_2011-07-19_000000_2011-07-19_235959.csv
tripdata_ids_2011-07-26_000000_2011-07-26_235959.csv
tripdata_ids_2011-08-02_000000_2011-08-02_235959.csv
tripdata_ids_2011-08-09_000000_2011-08-09_235959.csv
tripdata_ids_2011-08-16_000000_2011-08-16_235959.csv
tripdata_ids_2011-08-23_000000_2011-08-23_235959.csv
tripdata_ids_2011-08-30_000000_2011-08-30_235959.csv
tripdata_ids_2011-09-06_000000_2011-09-06_235959.csv
tripdata_ids_2011-09-13_000000_2011-09-13_235959.csv
tripdata_ids_2011-09-20_000000_2011-09-20_235959.csv
tripdata_ids_2011-09-27_000000_2011-09-27_235959.csv
tripdata_ids_2011-10-04_000000_2011-10-04_235959.csv
tripdata_ids_2011-10-11_000000_2011-10-11_235959.csv
tripdata_ids_2011-10-18_000000_2011-10-18_235959.csv
tripdata_ids_2011-10-25_000000_2011-10-25_235959.csv
tripdata_ids_2011-11-01_000000_2011-11-01_235959.csv
tripdata_ids_2011-11-08_000000_2011-11-08_235959.csv
tripdata_ids_2011-11-15_000000_2011-11-15_235959.csv
tripdata_ids_2011-11-22_000000_2011-11-22_235959.csv
tripdata_ids_2011-11-29_000000_2011-11-29_235959.csv
tripdata_ids_2011-12-06_000000_2011-12-06_235959.csv
tripdata_ids_2011-12-13_000000_2011-12-13_235959.csv
tripdata_ids_2011-12-20_000000_2011-12-20_235959.csv
tripdata_ids_2011-12-27_000000_2011-12-27_235959.csv

brenobeirigo/mod

Mobility-on-demand fleet management application

Installation

ADP config

Loading intances from files

Running the simulator

Running the service rate and fleet status graphs

Creating the environoment

Execution

Logging

Scenario configuration

Optimization methods

Demand

Execution

Memory profiling

Printing functools

Installing cython

MIP

Gurobi error codes

Gurobi method

Gurobi MIP focus

AMoD

Online and loaded data

Logged information

Methods

Reactive rebalancing

Tuning

Notebooks

Data

Map

Distance matrix

Node list

Node list and regional centers

First class distrubution

Pickup and delivery requests

Training data

Testing data