fix fitness and constraint plots

[jennmald]

• Fix the fitness plot issue, plots do not appear when they are supposed to
• add in the constraint plots using log_total_constraints

[jennmald]

@thomaswmorris I fixed up the broken plots and tried to follow your advice on the constraint plots. There are some constraint plots that look right and others that definitely don't. Can you review if those plots should even be included? Let me know what you think.

[thomaswmorris]

These already look a lot better. Which constraint plots are weird? Also, could you add the bad samples (maybe as a bunch of black x's) in the leftmost column?

[jennmald]

@thomaswmorris there are occasional plots that look like this:

where all the values in the dataset are the same.

Also what determines a "bad" point? Is there some threshold I should be computing?

[thomaswmorris]

"Bad" points are the nan values in agent.train_targets(). The constant values are normal, when there are no bad points the agent predicts all other points to be good with probability 1.

[jennmald]

@thomaswmorris if there are nan values in agent.train_targets() the _nanquantile function makes all points in the scatter plot not visible. There is no way I have found to determine which points correlate to np.nan values.
For example:
I made values the same as in the test but then injected a np.nan value at the end of the list for the last point. When the code hits:

values = obj._untransform(targets)
val_vmin, val_vmax = np.nanquantile(values, q=[0.01, 0.99])`
val_norm = (
            mpl.colors.LogNorm(val_vmin, val_vmax) if obj.transform == "log" else mpl.colors.Normalize(val_vmin, val_vmax)
        )

then the val_norm and everything after that is np.nan, so no values appear on the plot. Any advice on how to approach this?