Multi armed bandits, solving some issues, and more

[gAldeia]

This PR implements multi-armed bandits to learn sampling probabilities for terminals, operators, and variation operations.

It also implements an archive (Issue #58).

A new classification metric, average_precision_score, is also implemented (partially solving Issue #57), which was hard to do, especially because I needed it to work with lexicase selection based on single-test evaluations. Currently, lexicase is still based on log loss, but the survival, archive, and final individuals are picked with the average precision score.

It also implements complexity and linear complexity.

I added a feature to abide by feature types from pandas data frames on the Python side (Issue #56). However, this feature has caused me to notice some undesired conversions. Sometimes, I find myself using X.values to avoid copying datatypes from pandas.

There is a progress bar now (Issue #20) and different verbosity levels.

I fixed the mean label so it is no longer weighted, and there have also been many improvements in the logistic nodes for classification problems.

Some new unit tests (mainly in cpp) were implemented.

Performance is not my main concern right now, and I recognize several new TODOs and improvements, but I will focus on that later.

[lacava]

it should be based on the type of the SplitBest child argument. The threshold could be zero for a continuous variable, in which case you don't want to hide it.

[gAldeia]

I’ll add a new attribute to the split nodes to store the data type of the split feature. This seems to be the most straightforward way to quickly access the dtype of the split feature.

[lacava]

the type of the attribute for SplitBest is stored in arg_types[0].

wait, you already have the logic for SplitOn on line 48, can you just repeat that for SplitBest nodes?

    else if (Is<NodeType::SplitOn>(data.node_type)){
        if (data.arg_types.at(0) == DataType::ArrayB)
        {
            // booleans dont use thresholds (they are used directly as mask in split)
            return "If" + child_outputs;
        }

[lacava]

FYI there are three signatures for SplitBest nodes that are defined here, so you just have to look at the first arg type:

brush/src/program/signatures.h

Line 354 in ee44ac3

using type = std::tuple<