Solving Travelling Salesman Problems with Genetic Algorithms. Originally for SEM6120. The report for this assignment is located in the report folder. Documentation for this project is located in the doc/build/html folder.

Installation is easiest using the pip package manager. Python version 2.7.9+ automatically ship with pip. If you're using an older version of Python you can find the installation instructions here.

To install the program locally cd into the top level directory of the probject and run the following:

pip install -e .

All of the project dependancies should be installed automatically. If for some reason they are not, the full list of dependancies are:

• numpy
• scipy
• pandas
• scikit-learn
• matplotlib
• click

There are three major commands: generate, solve and tune.

The generate command will create a uniformly random TSP datasetand output it to the specified CSV file.

tspsolver generate dataset.csv

Optionally, it can also take a parameter specifying the number of cities to generate (default is 10):

tspsovler generate -n 30 dataset.csv

The solve command can be used to run the genetic algorithm and produce solutions to the problems. The command takes a JSON parameter file as an arguement. Optionally you can provide a dataset file (such as produced by the generate command) or specify a random number of cities to generate. When the command terminates, two plots are produced. One shows the min, mean, and max fitness across all generations, the second shows the best solution found over all datasets.

Example Parameter File:

{
    "num_epochs": 1000,
    "num_elites": 2,
    "generator": "SimplePopulationGenerator",
    "generator_population_size": 20,
    "selector": "RouletteWheelSelection",
    "selector_tournament_size": 5,
    "crossover": "OrderCrossover",
    "crossover_pcross": 0.9,
    "mutator": "InversionMutation",
    "mutator_pmutate": 0.1
}

Solving a TSP problem with a randomly generated TSP dataset:

tspsolver solve -n 20  params.json

Solving a TSP problem with an exisitng dataset:

tspsolver solve -f dataset.csv params.json

The final command can be used to run a range of parameter configurations over a number of different datasets. This can be useful to examine the effects of different parameter datasets. Each configuration is run on n different randomly generated datasets and the median result is take to represent the whoel. The results for all datasets are saved to a CSV file. The configuration that produced the best results is also saved to a JSON file.

This command takes a special parameter file that specifies ranges of parameters. An example is shown below:

{
    "num_epochs": [1000],
    "num_elites": [0],
    "generator": ["SimplePopulationGenerator"],
    "generator_population_size": [20],
    "selector": ["TournamentSelection"],
    "selector_tournament_size": [5],
    "crossover": ["OrderCrossover"],
    "crossover_pcross": [0.6, 0.7, 0.8, 0.9],
    "mutator": ["InversionMutation"],
    "mutator_pmutate": [0.01, 0.05, 0.1, 0.2]
}

This will run the genetic algorithm with a varying range of crossover and mutation probabilities. An example of running the tuning command is as follows:

tspsolver tune -d 5 -n 50 tuning_params.json results.csv best.json

In the above command the -d command specifies the number of datasets to generate for each parameter configuration. The -n flag specifies the number of random generated points to use for each dataset. tuning_params.json is the special parameter file with ranges. results.csv is the csv file created with all parameter results. best.json is the generated parameter file containing the parameters that produced the best run.