Plot how fitness changes over generations

[schrum2]

We need to have the ability to make graphs with generations on the x-axis and max fitness score on the y-axis. Every generation we evaluate the whole population, and save the max fitness, which we plot for that generation. In a text file, this means you have the max fitness values in a column next to a column of the generation numbers. From that it is easy to plot in a variety of ways.

NEAT-Python already has several "reporters" that are attached to the population in evolution.py. They are responsible for periodically printing information to the console, but they can probably be used for this too ... I'm just not sure how.

[alejmedinajr]

I think I found something similar to what needs to be in a text file from the documentation.

[schrum2]

Very good! I also found an example on their GitHub of how to use this. See: https://github.com/CodeReclaimers/neat-python/blob/5590604240793da311d9fb44671510139d9574f8/tests/test_simple_run.py Specifically:

    stats = neat.StatisticsReporter()

    stats.save()
    stats.save_genome_fitness(with_cross_validation=True)

This is a good first test, but we would probably want to add a filename as well.

According to this: https://github.com/CodeReclaimers/neat-python/blob/1d796adbb022cd3f84b4aecf678b26824e42b7f2/neat/statistics.py I think the default file name is fitness_history.csv, which we wouldn't always want

[alejmedinajr]

So I was able to get the fitness to save to csv file. The value for with_cross_validation has to be false because it is not supported yet (I was getting an error and when I looked at their github, there was a comment about how it had to be false). My next commit for this issue will focus more on the filename since I only focused on getting the thing to work and print to csv.

Attached is the fitness_history.csv produced from a type_target run where the desired number of blocks was 50 and the desired block was redstone block.

fitness_history.csv

[alejmedinajr]

I realized the plotting has to happen for fitness based evolution so I placed the csv saving inside an if condition that only hits when INTERACTIVE_EVOLUTION is false.

[schrum2]

We'll need a command line parameter for EXPERIMENT_NAME that is both the name of the csv file (or part of it) and the name of a results directory, kind of like MMNEAT

[alejmedinajr]

Before the default name for the produced csv file was "fitness_history.csv". Now it is dependent on some command line parameters, although I think one more can be added. Also now the csv file is saved in a sub directory related to its respective seed.

I will try to implement the visualize stuff as well.

[alejmedinajr]

Using visualize also work for visualize.plot_stats but got error I do not fully understand for visualize.plot_species. Also I had to run pip install graphviz in anaconda before I got the visualize to work. Not sure if this means the requirements.txt has to be updated to reflect this or if there is a better way of using graphviz.

[alejmedinajr]

Update, visualize.plot_species also works, I just forgot how it works.

[schrum2]

Update requirements.txt as well

[alejmedinajr]

Updated requirements.txt to have graphviz. I did not specify a specific version however.