Comp3620/Comp6320 Artificial Intelligence
Assignment 2: Constraint Programming
April 17, 2017
In this assignment, you will solve logistics problems using the constraint programming tool MiniZinc, see
http://www.minizinc.org/.
The Logistics Problem
Imagine a transport company which needs to send goods by road from a depot to a number of customers.
They have a fleet of trucks, of various sizes, some refrigerated and some not. The customers are far away
(hundreds of kilometres) so in one round of deliveries, each truck makes at most one journey out from the
depot, serving one or more customers, and returning to the depot. It is not necessary that all of the trucks
be used for this—in fact, usually it pays to use a fairly small subset of the available trucks. Trucks are
expensive things to run, so it is important to allocate the jobs to trucks in a way that reduces the overall
cost to the company: as you will see, a good solution to this problem can save thousands of dollars in a
single round of deliveries.
We consider goods of two sorts: chilled and ambient. Chilled goods can only be transported in refrigerated
trucks. Refrigerated trucks can also carry ambient-temperature goods, but they are more expensive to run
than non-refrigerated ones. In the version of the problem considered here, the number of customers is quite
small, but some of them have placed orders for more goods than will fit in one truck, so they must be visited
by more than one of the trucks. The goods are packed into containers, and each truck can carry a few of
these units of goods: larger trucks have greater capacity, but again they are more expensive to run than the
small ones.
In the setup of each problem instance, we know:
• Which trucks are available, how many units of goods they can carry, whether they are refrigerated or
not, and the cost per kilometre of running them;
• Which customers are to be served, and how many units of each goods type (chilled and ambient) they
have ordered;
• The distance by road between each pair of customers, and between each customer and the depot. In
some cases, the distance from A to B may not exactly equal the distance from B to A because of
(unspecified) details of the road network.
A solution to the logistics problem must specify for each truck the sequence of places it visits, starting
and finishing at the depot, and how many units of each of the two goods types it delivers to each customer
it visits. Of course, no truck may carry more than its capacity, and all customers get exactly what they
ordered. A truck that is not used in this particular delivery round just stays at the depot, thus making
a journey of length zero. The cost of each truck’s journey is the total distance (in kilometres) it travels
multiplied by the cost per kilometre, and the total cost of the operation is the sum of the journey costs for
all trucks in the fleet. An optimal solution is one that minimises this cost.
1
Minizinc
MiniZinc is not a programming language, but a language for specifying constraint satisfaction problems,
based on first order logic with a lot of built-in features and support for arithmetic. It comes with back-end
software which compiles the problem description into a low-level CSP which can be solved by a conventional
constraint-based reasoning system. This may be a finite domain solver, but it could be something else such as
a SAT solver or a mathematical programming tool. A central design feature of MiniZinc is that the logical
definition of the problem should be clearly separated from the program that solves it, so MiniZinc itself
encourages you to write solver-independent problem specifications, making it very suitable for our purposes.
Typically, the problem is specified in a file (with filename extension .mzn) in a way that enables different
instances of the problem to be encoded by supplying different data values. MiniZinc data files have filename
extension .dzn and will fix the values of parameters such as the number of trucks, the distances between
places and the orders placed by customers. This simple organisation of files enables all of the logistics
problems we consider here to be specified without changing the .mzn file at all.
For this assignment, you will need to design a logical description of the logistics problem, consistent
with the format for input of data and output of solutions described below. This means you must write the
constraints, and define the vocabulary they need. You must then run MiniZinc with your files as input, and
obtain solutions to the various problem instances.
Lab: Learning Minizinc
We’ll cover a fair number of the practical details of learning MiniZinc and getting MiniZinc to work in the
lab before your assignment is due. Go to the Lab Exercises directory and open the Lab CP students.pdf
file to see the exercises. They will guide you through a series of steps relevant to the techniques you will
need for this assignment.
The most helpful source of more comprehensive information about using MiniZinc is probably the tutorial
by Marriott and Stuckey, which is also included in the same directory. The material most useful for this
assignment is in sections 1 through 4—mostly in the part to the end of section 3.4, in fact. In section 4, only
the basics of using predicates and the alldifferent constraint are likely to be relevant.
For those who want the technical language spec, that’s also available online: http://www.minizinc.
org/doc-lib/minizinc-spec.pdf; for the rest of us, that’s not really necessary.
Assignment Specifics
Once you’ve completed the above lab exercises, you should be ready to start the assignment. The assignment
consists of a series of 4 steps each building on the previous one. These steps are to be completed within
directories Q1 to Q4, and are designed to progressively take you to the point where you have a complete
MiniZinc encoding of this problem and use it within a large neighborhood search (LNS).
For all problems, you will use the Finite Domain (FD) solver g12-fd. This is the default
solver which is called when running minizinc from the command line. However, if you use
the MiniZinc IDE, you may need to select this solver explicitly, as sometimes the IDE uses the
gecode solver by default. However, you are free to experiment with other solvers and report some of your
results in the .txt files mentioned below if you find they bring huge benefits in comparison to the FD solver
for your particular encoding.
Your grade will depend primarily on the correctness of your solution/encoding but also on its elegance and
efficiency, as well as your creativity (in particular for LNS), and the quality of your experimental observations.
2
Question 1: Allocating Deliveries to Trucks (30pts)
Problem description
Given the data on the capacities of trucks, and whether they are refrigerated or not, and given the orders
placed by some customers, find an allocation of goods to trucks which satisfies all customers orders without
exceeding the capacity of any truck and without allocating chilled goods to any non-refrigerated truck.
At this point, we are not concerned with the costs nor with the sequence in which the trucks will visit
customers. We also aren’t optimising anything, hence any solution satisfying the given constraints will do.
Files we give you
In the directory Q1 you will find:
• Logistics-Q1.mzn: the start of the MiniZinc encoding, merely declaring the parameters and types
required to read the data for this question.
• Logistics-Q1-4-3.dzn, Logistics-Q1-6-3.dzn, Logistics-Q1-12-4.dzn, Logistics-Q1-21-8.dzn:
4 data files corresponding to 4 problem instances of increasing difficulty. Logistics-Q1-x-y.dzn has
x trucks and y customers.
What you must produce
• In Logistics-Q1.mzn: add the MiniZinc variable declarations and constraints needed to solve the
problem for Q1, and a MiniZinc output item outputting the solution in the format described below.
Make sure you comment your file, as it will help us understand your intent, especially in case your
solution is incorrect overall.
• 4 files solution-Q1-4-3.csv, solution-Q1-6-3.csv, solution-Q1-12-4.csv, solution-Q1-21-8.csv:
each of these files should contain the solution (in the output format below) found by MiniZinc for the
respective problem instance. If you can’t solve one of the instances in the order of 1 minute1,
please create the file but leave the content empty. Note that your encoding may not be able to
solve the largest instance with the FD solver.
• 1 file report-Q1.txt: telling us, for each problem instance, whether it was solved at all and how long
it took to solve it. This is your chance to let us know about any problem you could solve but in times
that significantly exceed a minute, or about any substantially better results obtained with a different
solver. There is no precise format for this file. You might also want to report about other things that
aren’t evident from the final code or results, including any other encoding or technique your tried that
didn’t go so well. Please be reasonably concise.
Solution output format
For Q1, the solution output format is as follows:
nb_trucks, nb_customers (not including the depot)
[truck_id, customer_id, chilled_goods_units_delivered, ambient_goods_units_delivered]*
Hence the first line specifies the number of trucks and customers in the instance. Each other line spec-
ifies for a truck and a customer that trucks delivers to, how many units of chilled goods and how many
units of ambient goods that truck delivers to the customer. If the truck doesn’t deliver anything to a cus-
tomer, no corresponding line should be created. Here is an example of the beginning of a solution file for
solution-Q1-6-3.csv:
11 min 10 secs is OK, 2 min is pushing it, and 5 min is not OK.
3
6,3
1,1,0,4
3,3,7,0
4,2,4,0
4,3,3,0
etc
Please remove the ——— and ============ printed by MiniZinc. This will automatically happen if
you call MiniZinc as follows:
minizinc Logistics-Q1.mzn Logistics-Q1-6-3.dzn –soln-sep “” –search-complete-msg “” > solution-Q1-6-3.csv
MiniZinc has a rather unconventional way of specifying formatted output. To help you produce the re-
quired format, we stronly suggest that you have a look at some of the output examples in minizinc-tute.pdf,
especially those including the use of list comprehensions and the fix function. This function turns a variable
whose value has been determined (fixed) by the search, into a constant with this value. See in particular
p37-38. This is likely to be useful …
Question 2: Routing the Trucks (30pts)
Problem description
This is probably the trickest part of the assignment. The problem is now extended to include information
about the sequence of customers visited by each truck. Given the same data as for Question 1, find not
only the allocation of goods to trucks but also for each truck a sequence of places to be visited, starting and
finishing at the depot and in between visiting the customers to whom it makes deliveries. This sequence
should not visit the same customer twice, and it should not visit a customer to whom the truck does not
deliver anything.
Files we give you
In the directory Q2 you will find:
• Logistics-Q2.mzn: the start of the MiniZinc encoding, which is as for Q1 plus a new parameter
(times) giving the maximum number of time steps in a sequence. This is sufficient to (in the worst
case) visit all customers and return to the depot.
• Logistics-Q2-4-3.dzn, Logistics-Q2-6-3.dzn, Logistics-Q2-12-4.dzn, Logistics-Q2-21-8.dzn:
4 data files which are exactly the same as for Q1.
What you must produce in the directory Q2
• In Logistics-Q2.mzn: extend your Q1 encoding by adding the extra MiniZinc variable declarations
and constraints needed to solve the problem for Q2, and modifying the MiniZinc output item to match
the format described below.
• 4 files solution-Q2-4-3.csv, solution-Q2-6-3.csv, solution-Q2-12-4.csv, solution-Q2-21-8.csv:
as before create these files in the Q2 directory and print the solutions to the respective instances you
could solve in the order of 1 minute in the format below.
• 1 file report-Q2.txt: telling us, as before, for each problem, whether it was solved and how long it
took to solve it. Feel free to include any remark about what worked and what didn’t, and why.
4
Solution output format
This is similar to the format for Q1, but with the addition of an extra time index after the truck id to
indicate the time step in the sequence (1, 2 . . . ) at which the customer is visited:
nb_trucks, nb_customers (not including the depot)
[truck_id, time_id, customer_id, chilled_goods_units_delivered, ambient_goods_units_delivered]*
Hence the first line specifies the number of trucks and customers in the instance as for Q1. Each other line
specifies for a truck, the time index in the sequence at which it visits a given customer, the customer, and
how many units of chilled goods and how many units of ambient goods that truck delivers to the customer.
As before, if the truck doesn’t deliver anything to a customer, no corresponding line should be created, and
consequently no line should be included for the depot. Here is an example of the beginning of a solution file
for solution-Q2-6-3.csv:
6,3
1,1,1,0,4
3,1,3,7,0
4,1,2,4,0
4,2,3,3,0
etc
specifying, for instance, that truck 4 visits customer 2 at time step 1 and customer 3 at time step 2.
Question 3: Optimising (10pts)
Problem description
If you got that far, this question shouldn’t delay you too long. This part is as for Question 2, but this time
we require a solution that minimises the total cost. This means you must take into account the distance each
truck travels, and the cost running each particular truck to compute the total cost of the solution. Both the
distances between each pair of customers or the depot (in Km), and the running cost of each truck (in cents
per Km) are now given in the data files.
Files we give you
In the directory Q3 you will find:
• Logistics-Q3.mzn: the start of the MiniZinc encoding, which is as for Q2 but includes new arrays
to read the distance and cost data, as well as a new variable tot-cost representing the total solution
cost to be optimised.
• Logistics-Q3-4-3.dzn, Logistics-Q3-6-3.dzn, Logistics-Q3-12-4.dzn, Logistics-Q3-21-8.dzn:
4 data files now including the distance and cost data.
What you must produce
• In Logistics-Q3.mzn: extend your Q2 encoding by adding the extra MiniZinc variable declarations
and constraints needed to solve the problem for Q3, and slightly modifying the MiniZinc output item
to match the format described below.
• 4 files solution-Q3-4-3.csv, solution-Q3-6-3.csv, solution-Q3-12-4.csv, solution-Q3-21-8.csv:
as before create these files in the Q3 directory and print the optimal solutions to the respective in-
stances in the format below. Note that intermediate solutions printed by MiniZinc with the -a option
are not optimal. Create an empty file if you couldn’t solve the problem optimally in the
5
order of 1 minute. Optimising is notoriously more difficult than finding arbitrary solutions, hence
we do not expect that your encoding will be able to optimally solve Logistics-12-4 let alone the larger
problem with 21 trucks), but would be impressed if it does!
• 1 file report-Q3.txt: telling us, as before, for each problem, whether you could solve it optimally at
all and how long it took to solve it. Again here this is your chance to mention problems you could
solve but using significantly more time than 1 minute, other solvers tried, and anything else you think
we should know.
Solution output format
This is the same as for Q2 except that the first line now includes the total cost of the solution in dollars.
nb_trucks, nb_customers (not including the depot), tot_cost (in dollar)
[truck_id, time_id, customer_id, chilled_goods_units_delivered, ambient_goods_units_delivered]*
For example, the first line could be:
6,3,22678.70
Question 4: Scaling Up (30pts)
Problem description
If you are still with us, this is the fun part. For the first three questions, you were required only to write
MiniZinc encodings of the problems, and to note the limits on problem size that can be solved optimally, and
on problems that can be solved at all, by MiniZinc’s solver alone. In order to scale up to somewhat larger
problem instances, you will now implement large neighbourhood search (LNS) in Python, using MiniZinc
with its default FD solver to search each successive neighbourhood. Your LNS will start from a base solution
that we provide and will attempt to improve it.
For this question, you have some opportunity to try creative ideas as to how you might define the
neighbourhoods. You should also note the different effects of searching neighbourhoods for (locally) optimal
solutions and searching them merely for satisfying solutions. How well does your method scale up? How far
can you go?
Implementation Hints
There are many ways your Python program could interact with MiniZinc to implement LNS, and we do not
want to prescribe one; the only constraints we place are on what must be produced. For information, here
are three ideas we found useful in our own implementation:
• We sometimes wanted to search neighbourhoods for a (locally) optimal solution, and sometimes for
just any solution. To avoid duplicating the MiniZinc model, we removed the ‘solve’ item from the file
Logistics-Q4.mzn, and created two small files Logistics-sat.mzn or Logistics-opt.mzn, each of
which consists of just two lines, one being the instruction to include Logistics-Q4.mzn and the other
being the relevant ‘solve’ item (solve satisfy, or solve minimize tot cost).
• In order to specify which part of the solution to keep and which to remove, to define the neighbourhood,
we had our Python program write an additional datafile keep.dzn containing that information in the
form of array assignments. At each iteration of the LNS, we call MiniZinc with 3 parameters: one of
the mzn files (Logistics-sat.mzn or Logistics-opt.mzn), the data file for the original problem (for
instance Logistics-Q4-12-4.dzn), and keep.dzn.
6
• To assign values to part of an array, leaving another part blank, you can use the anonymous decision
variable “ ” (the underscore character). For example, given an array saying which of four people like,
which, we may write for instance
likes = array2d(person,person,
[true, _, false, true,
false, _, true, true,
false, _, false, true,
true, _, true, true ]);
to remove the specification of who likes person 2, while keeping the rest. We found this notation useful
for deleting just the parts of the solution we wanted to remove to create a neighbourhood.
Note that depending on the effectiveness of your encoding for Q3, it may be that LNS will not improve
much your overall results over vanilla MiniZinc, except perhaps on the largest instance (21 trucks). Or it
could be that it is able to find better solutions, even on smaller instances (e.g. 6 trucks).
Files we give you
• Logistics-LNS.py: a near-empty skeleton of your Python program implementing LNS. This skeleton
just enforces that the program takes as input the problem data file (e.g. Logistics-Q4-12-4.dzn)
and the solution file to improve (e.g. start-solution-12-4.dzn).
• Logistics-Q4-4-3.dzn, Logistics-Q4-6-3.dzn, Logistics-Q4-12-4.dzn, Logistics-Q4-21-8.dzn:
4 data files describing the problem instances, identical to those in Q3.
• start-solution-4-3.dzn, start-solution-6-3.dzn, start-solution-12-4.dzn, start-solution-21-8.dzn:
the initial solutions to improve on for the respective problem instances. The format for these files is
the solution output format for Q3.
• We also give you our Logistics-sat.mzn and Logistics-opt.mzn but you do not have to use them.
What you must return
• In Logistics-LNS.py, write code implementing LNS. This code must print to standard out the initial
solution to improve, and, as they are found by the LNS, any new solution found whose cost is better
than the current best solution.
• Any other file that your program needs to be able to run. E.g. our own program needs Logistics-Q4.mzn.
• best-solution-4-3.dzn, best-solution-6-3.dzn, best-solution-12-4.dzn, best-solution-21-8.dzn:
the best solution found by your LNS within 1 minute in the first three cases, and two minutes for
best-solution-21-8.dzn. Use the same solution output format as for Q3.
• 1 file report-Q4.txt: containing more details of how quickly your LNS actually converged to the best
solution on the smaller problems, and the results of any other experiments you ran on the large ones.
You should say how your neighbourhoods were defined—what you destroyed of each solution and what
parts of it you kept—and as before comment on what seems to work well or not so well on this logistics
problem.
Solution output format
As stated above, this is the same as for Q3.
7