Search
tsp

If any part of this notebook is used in your research, please cite with the reference found in README.md.


The Traveling Sales(man)(person) Problem — TSP

Integrating pysal/spaghetti and pulp for optimal routing

Author: James D. Gaboardi jgaboardi@gmail.com

This notebook provides a use case for:

  1. Introducing the TSP
  2. Declaration or a solution class and model parameters
  3. Solving the TSP for an optimal tour
%load_ext watermark
%watermark
2020-01-25T17:07:52-05:00

CPython 3.7.3
IPython 7.10.2

compiler   : Clang 9.0.0 (tags/RELEASE_900/final)
system     : Darwin
release    : 19.2.0
machine    : x86_64
processor  : i386
CPU cores  : 4
interpreter: 64bit

In addtion to the base spaghetti requirements (and their dependecies), this notebook requires installations of:

  • geopandas
    • $ conda install -c conda-forge geopandas
  • matplotlib
    • $ conda install matplotlib
  • pulp
    • $ conda install -c conda-forge pulp

import spaghetti
import geopandas
from libpysal import examples
import matplotlib
import numpy
import pulp
try:
    from IPython.display import set_matplotlib_formats

    set_matplotlib_formats("retina")
except ImportError:
    pass
%matplotlib inline
%watermark -w
%watermark -iv
watermark 2.0.2
geopandas  0.6.2
pulp       1.6.8
matplotlib 3.1.2
spaghetti  1.4.0
numpy      1.17.3


1 Introduction

Scenario

Detective George B. Königsberg thought he needed to visit 7 crimes scenes in one area of City X this afternoon in order to collect evidence. However, his lieutenant, Anna Nagurney just told him he needs to double that to 14. He really wants to wrap up early so he can get home to watch the 2012 mathematical thriller, Travelling Salesman by Timothy Lanzone), with his cat and dog, Euler and Hamilton. Therefore, he decides on calculating an optimal route so that he can visit all 14 crime scenes in one tour while covering the shortest distance. Det. Königsberg utilizes an integer linear programming formulation of the traveling salesperson problem (TSP) to find his best route.


Integer Linear Programming Formulation based on Miller, Tucker, and Zemlin (1960).

$\begin{array} \displaystyle \normalsize \textrm{Minimize} & \displaystyle \normalsize \sum_{0 \leq i \\ i \neq j}^n \sum_{j \leq n \\ j \neq i}^n c_{ij}x_{ij} & & & & \normalsize (1) \\ \normalsize \textrm{Subject To} & \displaystyle \normalsize \sum_{i=0}^n x_{ij}=1 & \normalsize j=1,...,n, & \normalsize j\neq i; & &\normalsize (2)\\ & \displaystyle \normalsize \sum_{j=0}^n x_{ij}=1 & \normalsize i=1,...,n, & \normalsize i\neq j; & &\normalsize (3) \\ & \displaystyle \normalsize u_i - u_j + p x_{ij} \leq p - 1 & \normalsize i=1,...,n, & \normalsize 1 \leq i \neq j \leq n; & &\normalsize (4) \\ & \normalsize x_{ij} \in \{0,1\} & \normalsize i=1,...,n, & \normalsize j=1,...,n; & &\normalsize (5)\\ & \normalsize u_{i} \in \mathbb{Z} & \normalsize i=1,...,n. & & &\normalsize (6)\\ \end{array}$

$\begin{array} \displaystyle \normalsize \textrm{Where} & \small x_{ij} & \small = & \small \begin{cases} 1, & \textrm{if node } i \textrm{ immediately precedes node } j \textrm{ in the tour}\\ 0, & \textrm{otherwise} \end{cases} &&&&\\ & \small c_{ij} & \small = & \small \textrm{distance matrix between all } i,j \textrm{ pairs} &&&& \\ & \small n & \small = & \small \textrm{the total number of nodes in the tour} &&&&\\ & \small i & \small = & \small \textrm{each potential origin node} &&&&\\ & \small j & \small = & \small \textrm{each potential destination node} &&&&\\ & \small u_i & \small = & \small \textrm{continuous, non-negative real numbers} &&&&\\ & \small p & \small = & \small \textrm{allowed visits prior to return (}n = p \textrm{ in this formulation)} &&&&\\ \end{array}$


References

  • Cummings, N. (2000) A brief History of the Travelling Salesman Problem. The Operational Research Society. Accessed: 01/2020.
  • Dantzig, G., Fulkerson, R., and Johnson, S. (1954) Solution of a Large-Scale Traveling-Salesman Problem. Journal of the Operational Research Society of America. 2(4)393-410.
  • Flood, Merrill M. (1956) The Traveling-Salesman Problem. 4(1)61-75.
  • Gass, S. I. and Assad, A. A. (2005) An Annotated Timeline of Operations Research: An Informal History. Springer US.
  • Miller, C. E., Tucker, A. W., and Zemlin, R. A. (1960) Integer Programming Formulation of Traveling Salesman Problems. Journal of Association for Computing Machinery. 7(4)326-329.
  • Miller, H. J. and Shaw, S.-L. (2001) Geographic Information Systems for Transportation: Principles and Applications. New York. Oxford University Press.
  • Nemhauser, G. L. and Wolsey, L. A. (1988) Integer and Combinatorial Optimization. John Wiley & Sons, Inc.

2. A model, data and parameters

Solution class

class MTZ_TSP:
    def __init__(self, nodes, cij, xij_tag="x_%s,%s", ui_tag="u_%s"):
        """Instantiate and solve the Traveling Salesperson Problem (TSP)
        based the formulation from Miller, Tucker, and Zemlin (1960).
        
        Parameters
        ----------
        nodes : geopandas.GeoSeries
            All nodes to be visited in the tour.
        cij : numpy.array
            All-to-all distance matrix for nodes.
        xij_tag : str
            Tour decision variable names within the model. Default is
            'x_%s,%s' where %s indicates string formatting.
        ui_tag : str
            Arbitrary real number decision variable names within the model.
            Default is 'u_%s' where %s indicates string formatting.
        
        Attributes
        ----------
        nodes : geopandas.GeoSeries
            See description in above.
        p : int
            The number of nodes in the set.   
        rp_0n : range
            Range of node IDs in ``nodes`` from 0,...,``p``.
        rp_1n : range
            Range of node IDs in ``nodes`` from 1,...,``p``.
        id : str
            Column name for ``nodes``.
        cij : numpy.array
            See description in above.
        xij_tag : str
            See description in above.
        ui_tag : str
            See description in above.
        tsp : pulp.LpProblem
            Integer Linear Programming problem instance.
        xij : numpy.array
            Binary tour decision variables (``pulp.LpVariable``).
        ui : numpy.array
            Continuous arbitrary real number decision variables
            (``pulp.LpVariable``).
        cycle_ods : dict
            Cycle origin-destination lookup keyed by origin with
            destination as the value.
        tour_pairs : list
            OD pairs comprising each abstract tour arc.
        """

        # all nodes to be visited and the distance matrix
        self.nodes, self.cij = nodes, cij
        # number of nodes in the set
        self.p = self.nodes.shape[0]
        # full and truncated range of nodes (p) in the set
        self.rp_0n, self.rp_1n = range(self.p), range(1, self.p)
        # column name for node IDs
        self.id = self.nodes.name
        # alpha tag for decision and dummy variable prefixes
        self.xij_tag, self.ui_tag = xij_tag, ui_tag

        # instantiate a model
        self.tsp = pulp.LpProblem("MTZ_TSP", pulp.LpMinimize)
        # create and set the tour decision variables
        self.tour_dvs()
        # create and set the arbitraty real number decision variables
        self.arn_dvs()
        # set the objective function
        self.objective_func()
        # node entry constraints
        self.entry_exit_constrs(entry=True)
        # node exit constraints
        self.entry_exit_constrs(entry=False)
        # subtour prevention constraints
        self.prevent_subtours()
        # solve
        self.tsp.solve()
        # origin-destination lookup
        self.get_decisions(display=display)
        # extract the sequence of nodes to construct the optimal tour
        self.construct_tour()

    def tour_dvs(self):
        """Create the tour decision variables - eq (5)."""

        def _name(_x):
            """Helper for naming variables"""
            return self.nodes[_x].split("_")[-1]

        xij = numpy.array(
            [
                [
                    pulp.LpVariable(self.xij_tag % (_name(i), _name(j)), cat="Binary")
                    for j in self.rp_0n
                ]
                for i in self.rp_0n
            ]
        )

        self.xij = xij

    def arn_dvs(self):
        """Create arbitrary real number decision variables - eq (6)."""
        ui = numpy.array(
            [pulp.LpVariable(self.ui_tag % (i), lowBound=0) for i in self.rp_0n]
        )

        self.ui = ui

    def objective_func(self):
        """Add the objective function - eq (1)."""
        self.tsp += pulp.lpSum(
            [
                self.cij[i, j] * self.xij[i, j]
                for i in self.rp_0n
                for j in self.rp_0n
                if i != j
            ]
        )

    def entry_exit_constrs(self, entry=True):
        """Add entry and exit constraints - eq (2) and (3)."""
        if entry:
            for i in self.rp_0n:
                self.tsp += (
                    pulp.lpSum([self.xij[i, j] for j in self.rp_0n if i != j]) == 1
                )
        # exit constraints
        else:
            for j in self.rp_0n:
                self.tsp += (
                    pulp.lpSum([self.xij[i, j] for i in self.rp_0n if i != j]) == 1
                )

    def prevent_subtours(self):
        """Add subtour prevention constraints - eq (4)."""
        for i in self.rp_1n:
            for j in self.rp_1n:
                if i != j:
                    self.tsp += (
                        self.ui[i] - self.ui[j] + self.p * self.xij[i, j] <= self.p - 1
                    )

    def get_decisions(self, display=True):
        """Fetch the selected decision variables."""
        cycle_ods = {}
        for var in self.tsp.variables():
            if var.name.startswith(self.ui_tag[0]):
                continue
            if var.varValue > 0:
                if display:
                    print("%s: %s" % (var.name, var.varValue))
                od = var.name.split("_")[-1]
                o, d = [int(tf) for tf in od.split(",")]
                cycle_ods[o] = d
        if display:
            print("Status: %s" % pulp.LpStatus[self.tsp.status])

        self.cycle_ods = cycle_ods

    def construct_tour(self):
        """Construct the tour."""
        tour_pairs = []
        for origin in self.rp_0n:
            tour_pairs.append([])
            try:
                tour_pairs[origin].append(next_origin)
                next_origin = self.cycle_ods[next_origin]
                tour_pairs[origin].append(next_origin)
            except NameError:
                next_origin = self.cycle_ods[origin]
                tour_pairs[origin].append(origin)
                tour_pairs[origin].append(next_origin)

        tour_pairs = {idx: sorted(cp) for idx, cp in enumerate(tour_pairs)}
        self.tour_pairs = tour_pairs

    def extract_tour(self, paths, id_col, leg_label="leg"):
        """Extract the tour (the legs in the journey) as a 
        ``geopandas.GeoDataFrame`` of ``shapely.geometry.LineString`` objects.
        
        Parameters
        ----------
        paths : geopandas.GeoDataFrame
            Shortest-path routes between all observations.
        id_col : str
            ID column name.
        leg_label : str
            Column name for the tour sequence. Default is 'leg'.
        
        Returns
        -------
        tour : geopandas.GeoDataFrame
            Optimal tour of ``self.nodes`` squenced by ``leg_label``.
        """
        paths[leg_label] = int
        # set label of journey leg for each OD pair.
        for leg, cp in self.tour_pairs.items():
            paths.loc[(paths[id_col] == tuple(cp)), leg_label] = leg
        # extract only paths in the tour
        tour = paths[paths[leg_label] != int].copy()
        tour.sort_values(by=[leg_label], inplace=True)

        return tour

Streets

streets = geopandas.read_file(examples.get_path("streets.shp"))
streets.crs = "epsg:2223"
streets = streets.to_crs("epsg:2762")

Crimes

all_crimes = geopandas.read_file(examples.get_path("crimes.shp"))
all_crimes.crs = "epsg:2223"
all_crimes = all_crimes.to_crs("epsg:2762")

Detective Königsberg's cases

numpy.random.seed(1960)
koenigsberg_cases = 7 * 2
subset_idx = numpy.random.choice(all_crimes.index, koenigsberg_cases, replace=False)
crimes_scenes = all_crimes[all_crimes.index.isin(subset_idx)].copy()

Instantiate a network object

ntw = spaghetti.Network(in_data=streets)
vertices, arcs = spaghetti.element_as_gdf(ntw, vertices=True, arcs=True)

Plot

base = arcs.plot(linewidth=3, alpha=0.25, color="k", zorder=0, figsize=(10, 10))
vertices.plot(ax=base, markersize=2, color="red", zorder=1)
all_crimes.plot(ax=base, markersize=5, color="k", zorder=2)
crimes_scenes.plot(ax=base, markersize=100, alpha=0.25, color="blue", zorder=2)
<matplotlib.axes._subplots.AxesSubplot at 0x1065734a8>

Associate Detective Königsberg's cases with the network and plot

ntw.snapobservations(crimes_scenes, "crime_scenes")
pp_obs = spaghetti.element_as_gdf(ntw, pp_name="crime_scenes")
pp_obs_snapped = spaghetti.element_as_gdf(ntw, pp_name="crime_scenes", snapped=True)
base = arcs.plot(linewidth=3, alpha=0.25, color="k", zorder=0, figsize=(10, 10))
vertices.plot(ax=base, markersize=5, color="r", zorder=1)
pp_obs.plot(ax=base, markersize=20, color="k", zorder=2)
pp_obs_snapped.plot(ax=base, markersize=20, marker="x", color="k", zorder=2)
<matplotlib.axes._subplots.AxesSubplot at 0x10a2227b8>

Calculate distance matrix while generating shortest path trees

d2d_dist, tree = ntw.allneighbordistances("crime_scenes", gen_tree=True)
d2d_dist[:3, :3]
array([[          nan,  877.47164862, 1012.61083447],
       [ 877.47164862,           nan,  688.1148453 ],
       [1012.61083447,  688.1148453 ,           nan]])
list(tree.items())[:4], list(tree.items())[-4:]
([((0, 1), (164, 158)),
  ((0, 2), (164, 142)),
  ((0, 3), (164, 197)),
  ((0, 4), (164, 147))],
 [((10, 13), (72, 98)),
  ((11, 12), (26, 29)),
  ((11, 13), (26, 85)),
  ((12, 13), (30, 85))])

3. The Travling Salesperson Problem

Create decision variables for the crime scene locations

pp_obs["dv"] = pp_obs["id"].apply(lambda _id: "x_%s" % _id)
pp_obs
id geometry dv
0 0 POINT (221651.779 266962.433) x_0
1 1 POINT (220774.565 266967.614) x_1
2 2 POINT (220962.017 267280.644) x_2
3 3 POINT (220772.431 267386.410) x_3
4 4 POINT (220796.815 267545.820) x_4
5 5 POINT (221500.903 267697.915) x_5
6 6 POINT (220474.032 267730.529) x_6
7 7 POINT (221677.382 267761.314) x_7
8 8 POINT (221154.041 267919.810) x_8
9 9 POINT (221852.947 268049.045) x_9
10 10 POINT (220775.174 268149.629) x_10
11 11 POINT (220472.813 268197.482) x_11
12 12 POINT (220514.875 268478.203) x_12
13 13 POINT (222083.681 268590.979) x_13

Solve the TSP

mtz_tsp = MTZ_TSP(pp_obs["dv"], d2d_dist)
x_0,1: 1.0
x_1,2: 1.0
x_10,8: 1.0
x_11,12: 1.0
x_12,10: 1.0
x_13,7: 1.0
x_2,3: 1.0
x_3,4: 1.0
x_4,6: 1.0
x_5,0: 1.0
x_6,11: 1.0
x_7,5: 1.0
x_8,9: 1.0
x_9,13: 1.0
Status: Optimal

Extract all network shortest paths

paths = ntw.shortest_paths(tree, "crime_scenes")
paths_gdf = spaghetti.element_as_gdf(ntw, routes=paths)
paths_gdf.head()
id geometry
0 (0, 1) LINESTRING (221652.388 266992.490, 221523.237 ...
1 (0, 2) LINESTRING (221652.388 266992.490, 221523.237 ...
2 (0, 3) LINESTRING (221652.388 266992.490, 221523.237 ...
3 (0, 4) LINESTRING (221652.388 266992.490, 221523.237 ...
4 (0, 5) LINESTRING (221652.388 266992.490, 221523.237 ...

Extract the tour

tour = mtz_tsp.extract_tour(paths_gdf, "id")
tour.head()
id geometry leg
0 (0, 1) LINESTRING (221652.388 266992.490, 221523.237 ... 0
13 (1, 2) LINESTRING (220775.021 267000.303, 220807.418 ... 1
25 (2, 3) LINESTRING (220961.989 267315.817, 220853.037 ... 2
36 (3, 4) LINESTRING (220807.120 267386.436, 220806.992 ... 3
47 (4, 6) LINESTRING (220796.691 267553.295, 220695.386 ... 4

Define label helper functions and plot Det. Königsberg's optimal tour

def tour_labels(t, b):
    """Label each leg of the tour."""

    def _lab_loc(_x):
        """Helper for labeling location."""
        return _x.geometry.interpolate(0.5, normalized=True).coords[0]

    kws = {"size": 20, "ha": "center", "va": "bottom", "weight": "bold"}
    t.apply(lambda x: b.annotate(s=x.leg, xy=_lab_loc(x), **kws), axis=1)


def obs_labels(o, b):
    """Label each point pattern observation."""

    def _lab_loc(_x):
        """Helper for labeling observations."""
        return _x.geometry.coords[0]

    kws = {"size": 14, "ha": "left", "va": "bottom", "style": "oblique", "color": "k"}
    o.apply(lambda x: b.annotate(s=x.id, xy=_lab_loc(x), **kws), axis=1)
base = arcs.plot(alpha=0.2, linewidth=1, color="k", figsize=(10, 10), zorder=0)
tour.plot(ax=base, column="id", cmap="tab20", alpha=0.50, linewidth=10, zorder=1)
vertices.plot(ax=base, markersize=1, color="r", zorder=2)
pp_obs.plot(ax=base, markersize=20, color="k", zorder=3)
pp_obs_snapped.plot(ax=base, markersize=20, color="k", marker="x", zorder=2)
# tour leg labels
tour_labels(tour, base)
# crime scene labels
obs_labels(pp_obs, base)