Experiments with Frequency Fitness Assignment based Algorithms on the Traveling Salesperson Problem

1. Introduction
In this archive, we provide the implementation and experimental results of eight different algorithms to solve Traveling Salesperson Problem (TSP) instances from TSPLIB.
A TSP is defined by a fully-connected weighted graph of n cities. The goal is to find the overall shortest tour that visits each cities exactly once and returns to its starting point. The TSP is NP-hard. We consider 56 symmetric instances from the well-known TSPLIB.
Solutions in our work are stored in the path representation, where such a tour is encoded as a permutation x of the numbers 1 to n, each identifying a city. If a city appears at index j in the permutation x, then it will be the jth city to be visited. This means that a tour x will pass the following edges: (x[1], x[2]), (x[2], x[3]), (x[3], x[4]), … (x[n-1], x[n]), (x[n], x[1]).
2. Directory Structure
This dataset is split into multiple separate tar.xz archives. These can be unpacked in the same folder and will produce the directory structure described below. Each archive contains this note and the license information, but apart from that, there is no redundancy.
This archive contains the following directories:

source contains the Python source codes needed to run the experiment.

moptipy-main is a local copy of the moptipy package used for our experiment.
tsplib contains the TSPLIB data. This includes the instances used in our experiments as files in text format with suffix .tsp. If an optimal tour is given by TSPLib, it is stored in a text format file with suffix .opt.tour and name prefix identical to the instance file. In other words, the file eil51.tsp contains the TSP instance eil51 and the file eil51.opt.tour contains the corresponding optimal tour. Both the TSP instances and optimal tours can be downloaded from http://comopt.ifi.uni-heidelberg.de/software/TSPLIB95/tsp/. We also include the documentation of TSPLIB in file tsp95.pdf documenting them. We further include the TSPLIB FAQ both as HTML and PDF file (tsplib_faq.html and tsplib_faq.pdf) and the list of known optimal tour lengths as HTML and PDF file (optimal_tour_lengths_of_symmetric_tsps.html, optimal_tour_lengths_of_symmetric_tsps.pdf). Notice that, while the TSP instances we used are Euclidean, all distances are converted to integers as prescribed by the documentation.
results is the directory with the log files. Each log file contains information of one run, i.e., one execution of one algorithm on one problem instance. All improving moves of a run as well as the final solution are stored in the log file. The direct sub-folders results represent the algorithms and contain one folder per TSP instance, which, in turn, contain the log files.
evaluation is a folder with the extracted evaluation and figures
evaluation_edited is a folder with evaluation figures slightly edited for better visual appeal (but obviously without changing any result / scientific content)
evaluator is a folder with a Python script main.py that generates all the files in evaluation from the data it finds in results. It requires the moptipy package being installed for running in the version given in requirements.txt.
3. Algorithms
The (1+1) EA is the most basic evolutionary algorithm and also be considered as a randomized local search. It starts with one random solution/permutation xc and computes its length yc=f(xc). In each iteration, it applies a unary search operator op to obtain a new tour xn=op(xc) and computes its length yn=f(xn). If yn<=yc, then it will accept the new tour and set xn=xn and yc=yn. The results of this algorithm are given in folder results/ea_revn.
FFA is a fitness assignment process that takes place before this last step in the EA. We integrate FFA into the (1+1) EA and obtain the (1+1) FEA. This algorithm uses an additional table H which counts, for any tour length y, how often it has been seen during the search so far. After the new tour xn is created and its objective value yn is computed, the (1+1) FEA sets H[yc] = H[yc] + 1 and H[yn] = H[yn] + 1. It will accept xn if and only if H[yn] <= H[yc] and, only in this case, set xn=xn and yc=yn. The results of this algorithm are given in folder results/fea_revn.
SA is the classical simulated annealing algorithm. In our experiment, it will accept the new solution xn with probability P. If the new solution is better, the acceptance probability P is 1. For worse solutions, the probability is between 0 and 1, i.e., sometimes, worse solution are also accepted. This algorithm has a temperature cooling schedule. It starts at an initial temperature and over time, the temperature decreases. The probability P of accepting the worse solution depends on the temperature and decreases as well. The results of this algorithm are given in folder results/sa_revn.
An FFA-based version of SA uses the frequency fitness instead of the objective values in all acceptance decisions. The results of this algorithm are given in folder results/fsa_revn.
EAFEA(A) is a hybrid which alternates between the EA and the FEA and copies a solution from the FEA to the EA if it has an entirely new objective value, i.e., if H[yn] = 1. The results of this algorithm are given in folder results/eafea2_revn.
SAFEA(A) is a hybrid which alternates between the SA and the FEA and copies a solution from the FEA to the SA if it has an entirely new objective value, i.e., if H[yn] = 1. The results of this algorithm are given in folder results/safea2_revn.
EAFEA(B) is a hybrid which alternates between the EA and the FEA and copies a solution from the FEA to the EA part if it has a better objective value. The results of this algorithm are given in folder results/eafea_revn.
SAFEA(B) is a hybrid which alternates between the SA and the FEA and copies a solution from the FEA to the SA part if it has a better objective value. The results of this algorithm are given in folder results/safea_revn.
We apply all algorithms with the same unary operator reverse, which reverses a randomly chosen subsequence of the tour. This operator is also often called a "2-opt move". It has the advantage that the new objective value of a new solution can be computed in O(1) if the objective value of the solution from which it is derived is known.
4. How to Run the Experiment
First, you need to make sure to have all the dependencies installed that this program requires. You can do this by executing the following command in the terminal:
pip install matplotlib numba numpy psutil scikit-learn moptipy moptipyapps
Now enter the source directory, i.e., the directory containing the run.py file, in your terminal. Depending on your system configuration and whether you run Windows or Linux, you can start the program with one of the commands below. (If running the first command returns with an error, just try the next one in the list.)

python3 -m run
python -m run
python run.py
python3 run.py
Then the experiment will run. It will automatically create a sub-folder results in source and place all log files that are generated into it. Be careful: The experiment will take a long time. However, if you have multiple CPUs, you can simply start several instances of this program in independent terminals. Each instance will then conduct different runs. This also works if this folder is shared over the network, in which case you can run multiple processes on multiple PCs.
Side note: This experiment uses the moptipy package for implementing its algorithms, running the experiments, and gathering their results. If you want to install moptipy on your system instead of using the version supplied here, you can install it via pip install moptipy. It also uses moptipyapps to load some data.
5. Literature

Frequency Fitness Assignment (FFA):

Thomas Weise, Zhize Wu, Xinlu Li, Yan Chen, and Jörg Lässig. Frequency Fitness Assignment: Optimization without Bias for Good Solutions can be Efficient. IEEE Transactions on Evolutionary Computation (TEVC). 2022. Early Access. doi:10.1109/TEVC.2022.3191698.
Thomas Weise, Zhize Wu, Xinlu Li, and Yan Chen. Frequency Fitness Assignment: Making Optimization Algorithms Invariant under Bijective Transformations of the Objective Function Value. IEEE Transactions on Evolutionary Computation 25(2):307–319. April 2021. Preprint available at arXiv:2001.01416v5 [cs.NE] 15 Oct 2020. doi:10.1109/TEVC.2020.3032090. Experimental results and source code are available at doi:10.5281/zenodo.3899474.
Tianyu Liang, Zhize Wu, Jörg Lässig, Daan van den Berg, and Thomas Weise. Solving the Traveling Salesperson Problem using Frequency Fitness Assignment. In Hisao Ishibuchi, Chee-Keong Kwoh, Ah-Hwee Tan, Dipti Srinivasan, Chunyan Miao, Anupam Trivedi, and Keeley A. Crockett, editors, Proceedings of the IEEE Symposium on Foundations of Computational Intelligence (IEEE FOCI'22), part of the IEEE Symposium Series on Computational Intelligence (SSCI 2022). December 4–7, 2022, Singapore, pages 360–367. IEEE. doi:10.1109/SSCI51031.2022.10022296.
Thomas Weise, Mingxu Wan, Ke Tang, Pu Wang, Alexandre Devert, and Xin Yao. Frequency Fitness Assignment. IEEE Transactions on Evolutionary Computation (IEEE-EC) 18(2):226-243, April 2014. doi:10.1109/TEVC.2013.2251885.
Thomas Weise, Xinlu Li, Yan Chen, and Zhize Wu. Solving Job Shop Scheduling Problems Without Using a Bias for Good Solutions. In Genetic and Evolutionary Computation Conference Companion (GECCO'21 Companion), July 10-14, 2021, Lille, France. ACM, New York, NY, USA. ISBN 978-1-4503-8351-6. doi:10.1145/3449726.3463124.
Thomas Weise, Yan Chen, Xinlu Li, and Zhize Wu. Selecting a diverse set of benchmark instances from a tunable model problem for black-box discrete optimization algorithms. Applied Soft Computing Journal (ASOC), 92:106269, June 2020. doi:10.1016/j.asoc.2020.106269.
Thomas Weise, Mingxu Wan, Ke Tang, and Xin Yao. Evolving Exact Integer Algorithms with Genetic Programming. In Proceedings of the IEEE Congress on Evolutionary Computation (CEC'14), Proceedings of the 2014 World Congress on Computational Intelligence (WCCI'14), pages 1816-1823, Beijing, China, July 6-11, 2014. Los Alamitos, CA, USA: IEEE Computer Society Press. ISBN: 978-1-4799-1488-3. doi:10.1109/CEC.2014.6900292.
Traveling Salesperson Problem (TSP):

Pedro Larrañaga, Cindy M. H. Kuijpers, Roberto H. Murga, I. Inza, and S. Dizdarevic. Genetic Algorithms for the Travelling Salesman Problem: A Review of Representations and Operators. Artificial Intelligence Review, 13(2):129–170, April 1999. Kluwer Academic Publishers, The Netherlands. doi:10.1023/A:1006529012972.
Gerhard Reinelt. TSPLIB — A Traveling Salesman Problem Library. ORSA Journal on Computing 3(4):376-384. 1991. http://comopt.ifi.uni-heidelberg.de/software/TSPLIB95/.
Gerhard Reinelt. TSPLIB95. 1995. Heidelberg, Germany: Universität Heidelberg, Institut für Angewandte Mathematik. http://comopt.ifi.uni-heidelberg.de/software/TSPLIB95/tsp95.pdf.
Thomas Weise, Raymond Chiong, Ke Tang, Jörg Lässig, Shigeyoshi Tsutsui, Wenxiang Chen, Zbigniew Michalewicz, and Xin Yao. Benchmarking Optimization Algorithms: An Open Source Framework for the Traveling Salesman Problem. IEEE Computational Intelligence Magazine (CIM) 9(3):40-52, August 2014. doi:10.1109/MCI.2014.2326101.
Eugene Leighton Lawler, Jan Karel Lenstra, Alexander Hendrik George Rinnooy Kan, and David B. Shmoys. The Traveling Salesman Problem: A Guided Tour of Combinatorial Optimization. Wiley Interscience. 1985.
David Lee Applegate, Robert E. Bixby, Vasek Chvatal, and William John Cook. The Traveling Salesman Problem: A Computational Study. Princeton University Press. 2007.
Gregory Z. Gutin and Abraham P. Punnen, editors. The Traveling Salesman Problem and its Variations. Volume 12 of Combinatorial Optimization. Kluwer Academic Publishers. 2002. doi:10.1007/b101971.
Software:

The Metaheuristic Optimization in Python Package moptipy
6. License
The files in this repository are under the Creative Commons Attribution 4.0 International, with the exception of the files of TSPLIB in directory source/tsplib, which are under copyright of their respective owner (we believe that they are in the public domain, as they are provided by many sources, included in many software packages under various open source licenses, and on many websites). The license is contained as file LICENSE.txt in this archive.
7. Contact
If you have any questions or suggestions, please contact
Mr. Tianyu LIANG (梁天宇) of the Institute of Applied Optimization (应用优化研究所, IAO) of the School of Artificial Intelligence and Big Data (人工智能与大数据学院) at Hefei University (合肥学院) in Hefei, Anhui, China (中国安徽省合肥市) via email to [e-mail].