Artificial Intelligence Methods (COMP2051 or AE2AIM)

Prof. Ruibin Bai Spring 2024

Coursework: Perturbative hyper-heuristic for Bin Packing Problem

1. Introduction
Bin packing is one of the most studied combinatorial optimisation problems and has
applications in logistics, space planning, production, cloud computing, etc. Bin packing is
proven to be NP-Hard and the actual difficulties depend on both the size of the problem (i.e.
the total number of items to be packed) and other factors like the distribution of item sizes in
relation to the bin size as well as the number of distinct item sizes (different items may have a
same size).

In this coursework, you are asked to write a C/C++/Python program to solve this problem
using a perturbative hyper-heuristic method. In addition to submitting source code, a
written report (no more than 2000 words and 6 pages) is required to describe your algorithm
(see Section 4 for detailed requirements). Both your program and report must be completed
independently by yourself. The submitted documents must successfully pass a plagiarism
checker before they can be marked. Once a plagiarism case is established, the academic
misconduct policies shall be applied strictly.

This coursework carries 45% of the module marks.

2. Bin Packing Problem (BPP)

Given a set of n items, each item j has a size of aj, BPP aims to pack all items in the
minimum number of identical sized bins without violating the capacity of bins (V). The
problem can be mathematically formulated as follow:

Artificial Intelligence Methods (COMP2051 or AE2AIM) Coursework Ver1.0 1

This mathematical formulation is generally NOT solvable by existing integer programming
solvers like CPlex, Gurobi, LPSolve, especially when the number of items n is large. The
solution space of bin packing problem is characterised by its huge size and plateau-like that
makes it very challenging for traditional neighbourhood search methods. In order to
consistently solve the problem with good quality solutions, metaheuristics and hyper-
heuristics are used, which is the task of this coursework.

3. Problem instances
Over the years, a large number of BPP instances have been introduced by various research.
See https://www.euro-online.org/websites/esicup/data-sets/ for a collection of different bin
packing problem. In this coursework, we shall provide 3 instances files (binpack1.txt,
binpack3.txt and binpack11.txt), respectively representing easy, medium and hard instances.
From which 10 instances shall be selected for testing and evaluation of your algorithm in
marking. For each test instance, only 1 run is executed, and its objective value is used for
marking the performance component (see Section 5).

4. Experiments conditions and submission requirements

The following requirements should be satisfied by your program:
(1) You are required to submit two files exactly. The first file should contain all your
program source codes. The second file is a coursework report. Please do NOT
compress the files.
(2) Your source code should adopt a clean structure and be properly commented.

Artificial Intelligence Methods (COMP2051 or AE2AIM) Coursework Ver1.0 2

 (3) Your report should include the followings:
• The main components of the algorithm, including solution encoding, fitness
function, list of low-level heuristics as well as considerations regarding the
intensification and diversification mechanisms. (12 marks).
• Statistical results (avg, best, worst of 5 runs) of the algorithm for all the problem
instances, in comparison with the best published results (i.e. the absolute gap to
the best results). Note that although your report should include results for 5 runs
but your final submission should only have one single run for each instance (i.e.
if you use the sketch code from the lab, set global variable NUM_OF_RUNS=1
before you submit the code). (3 marks)
• A short discussion/reflection on results and performance of the algorithm. (5
marks)

(4) Name your program file after your student id. For example, if your student number
is 2019560, name your program as 2019560.c (or 2019560.cpp, or 2019560.py).
(5) Your program should compile and run without errors on either CSLinux Server or a
computer in the IAMET406. Therefore, please fully tested before submission. You
may use one of the following commands (assuming your student id is 2019560 and
your program is named after your id):
gcc -std=c99 -lm 2019560.c -o 2019560
or
g++ -std=c++11 -lm 2019560.cpp -o 2019560
For Python programs, this second can be skipped.
(6) After compilation, your program should be executable using the following
command:
./2019560 -s data_fle -o solution_file -t max_time
where 2019560 is the executable file of your program, data_file is one of
problem instance files specified in Section 3. max_time is the maximum time
permitted for a single run of your algorithm. In this coursework, maximum of 30
seconds is permitted. soluton_file is the file for output the best solutions by
your algorithm. The format should be as follows:
# of problems
Instance_id1
obj= objective_value abs_gap
item_indx in bin0
item_indx in bin1
… …
Instance_id2
obj= objective_value abs_gap
item_indx in bin0

Artificial Intelligence Methods (COMP2051 or AE2AIM) Coursework Ver1.0 3

 item_indx in bin1
… …

An example solution file for problem data file “binpack1.txt” is available on

moodle.

For submissions using Python, the compilation and running are combined in one
command as follows:
python 2019560.py -s data_fle -o solution_file -t max_time

(7) The solution file output in (6) by your algorithm (solution_file) is expected to
pass a solution checking test successfully using the following command on
CSLInux:
./bpp_checker -s problem_file -c solution_file
where problem_file is one of problem data files in Section 3. If your solution file
format is correct, you should get a command line message similar to: “Your total score
out of 20 instances is: 80." If the solutions are infeasible for some instances, you would
get error messages.

The solution checker can be downloaded from moodle page. It is runnable only on
CSLinux.
(8) Your algorithm should run only ONCE for each problem instance and each run
should take no more than 30 seconds.
(9) Please carefully check the memory management in your program and test your
algorithm with a full run on CSLinux (i.e. running multiple instances in one go). In
the past, some submitted programs can run for 1-2 instances but then crashed
because of out-of-memory error. This, if happens, will greatly affect your score.
(10) You must strictly follow policies and regulations related to Plagiarism. You are
prohibited from using recent AI tools like ChatGPT/GPT-4 or other similar large
language models (LLMs). Once a case is established, it will be treated as a
plagiarism case and relevant policies and penalties shall be applied.

Artificial Intelligence Methods (COMP2051 or AE2AIM) Coursework Ver1.0 4

5. Marking criteria
• The quality of the experimental results (20 marks). Your algorithm shall be tested for
a file containing 10 instances chosen from the provided set of instances. The
performance of your algorithm is evaluated by computing the absolute gap with the
best known results using
𝑎𝑏𝑠_𝑔𝑎𝑝 = 𝑦𝑜𝑢𝑟_𝑎𝑣𝑒𝑟𝑎𝑔𝑒_𝑜𝑏𝑗𝑒𝑐𝑡𝑖𝑣𝑒 − 𝑏𝑒𝑠𝑡_𝑘𝑛𝑜𝑤𝑛_𝑜𝑏𝑗𝑒𝑐𝑡𝑖𝑣𝑒

Criteria Mark
New best results! Bonus: 2 extra marks for
abs_gap < 0
each new best result.
abs_gap <= 0 2 marks per instance

0<abs_gap <=1 1.5 marks per instance

1<abs_gap<=2 1 mark per instance

2< abs_gap <=3 0.5 mark per instance

• abs_gap >4 or
• infeasible solution or
0 mark
• fail to output solution
within required time limit

• The quality of codes, including organisation of the functions/methods, naming

conventions and clarity and succinctness of the comments (5 marks)
• Report (20 marks)

6. Submission deadline
3rd May 2024, 4pm Beijing Time

Standard penalties are applied for late submissions.

7. How to submit
Submit via Moodle.
8. Practical Hints
• Solution encoding for bin packing is slightly more challenging compared with
knapsack program because both the number of bins to be used and the number of
items to be packed in each bin are parts of decisions to be optimised. Therefore, the

Artificial Intelligence Methods (COMP2051 or AE2AIM) Coursework Ver1.0 5

 data structure that is used to hold the packing information cannot be implemented via
fixed-size arrays. You may consider to use vector from C++ STL (standard template
library) which requires you to include <vector.h> as header file. If you prefer C style
without classes, the following data type would be also acceptable:

struct bin_struct {
std::vector<item_struct> packed_items;
int cap_left;
};
struct solution_struct {
struct problem_struct* prob; //maintain a shallow copy of problem data
float objective;
int feasibility; //indicate the feasibility of the solution
std::vector<bin_struct> bins;
};

In this way, you could open/close bins and at the same time to add/remove items for a
specific bin through API functions provided by the vector library.
• The search space of bin packing problem has a lot of plateaus that make the problem
extremely difficult for simple neighbourhood methods. Therefore, multiple low-level
heuristics are suggested within a perturbative hyper-heuristic method. You are free to
select any of the perturbative hyper-heuristic methods described in
(https://link.springer.com/article/10.1007/s10288-011-0182-8), as well as some of the
more recent ones
(https://www.sciencedirect.com/science/article/pii/S0377221719306526).
• Your algorithm must be runnable on CSLinux and/or computers on IAMET406.
Therefore, you are not permitted to use external libraries designed specifically for
optimisation.

Artificial Intelligence Methods (COMP2051 or AE2AIM) Coursework Ver1.0 6