A small genetic algorithm developed in C with the objective of solving the Travelling Salesman Problem.
The genetic.c
file contains some explanation of how the program works.
Instead, progetto_algoritmi.pdf
file contains a detailed explanation of the code, the algorithms used and an analisys of the spatial and time complexity (in italian).
To compile and execute, use the commands:
gcc -o genetic.exe genetic.c -lm
./genetic.exe
If DEBUG is disabled, you will see only a couple of generations with the scores.
To draw the result of the execution, after the program ended use processing to run graphic.pde
.
This script will replicate the various generations and show graphically how the generations improved.
Here's a comparison between the first and last generation on a 750x750 map, with 50 cities, 256 individuals (every "individual" is a solution, e.g. a path).
- Choosing better hyper parameters, for a given problem, can reduce the required number of iterations and avoid local minimums
- Changing the cycle crossover (CX) to a better crossover function. The CX is one of the slowest crossovers for genetic algorithms (as shown here), but it is the easiest to implement. Other crossovers also have a better complexity.
- Changing the way in which the population is extracted for the reproduction can improve a lot when it comes to execution times. Extracting an index from the repartition array takes O(n) time, but using something like the Vose's Alias Method can reduce this time to O(1). This extraction is done for every member of the population array, so this improvement would make the program a lot faster, allowing for more iterations.
- rand() is used all over the code, however, it's a low quality random number generator. Moreover, using it with %SOMENUMBER creates an additional bias. C++'s <random> should be better, as well as Linux's PRNG via getrandom().
- Many other small improvements can be made, here only the most important ones were explained.
This program was the final project of the course "Algorithms and data structures" at Università Politecnica delle Marche (A.Y. 2016-2017) and was developed by Alessandro Cingolani, Giacomo Astolfi, Orazio Edoardo, Cristian Federiconi, Federica Massacci and Luca Luzi.