Bidirectional Active Processing (BAP) Implementation

Reimplementation of the BAP algorithm from Bidirectional Active Processing.md as specified in our Electronics paper available both online and locally. Single-file implementations in Python and Julia with TOML configuration.

Requirements

Python

• Python 3.11
• numpy, pandas, scikit-learn
• tomli (for Python < 3.11)

Use the project venv when present (dependencies are installed there):

source .venv311/bin/activate   # macOS / Linux
python bap.py -c examples/configs/config_iris.toml

Or run without activating:

.venv311/bin/python bap.py -c examples/configs/config_iris.toml

To create or refresh the venv:

python3.11 -m venv .venv311
.venv311/bin/python -m pip install -r requirements.txt

Configs point at CSVs under computing/machine learning datasets (not copies inside other app repos).

Julia

• Julia 1.8+
• Packages: CSV, DataFrames, ScikitLearn, TOML
  (ScikitLearn requires Python with scikit-learn)

julia -e 'using Pkg; Pkg.activate("./julia"); Pkg.instantiate()'

(Run from the repository root.)

Quick Start

Single CSV (train/test split)

Example: Fisher Iris dataset

# Python (after: source .venv311/bin/activate)
python bap.py -c examples/configs/config_iris.toml

# Or with flags
python bap.py --train "../../machine learning datasets/default/fisher_iris.csv" --testing split --split 0.8,0.2 --classifier dt -t 0.95 -n 10 -m 5

Separate train and test CSVs

Example: MNIST

# Python
python bap.py -c examples/configs/config_mnist.toml

# Or with flags
python bap.py --train "../../machine learning datasets/default/mnist_train.csv" --test "../../machine learning datasets/default/mnist_test.csv" --testing fixed --classifier knn --k 5 -n 5 -m 100

Julia

julia julia/bap.jl -c examples/configs/config_iris.toml
julia julia/bap.jl -c examples/configs/config_mnist.toml
julia julia/bap.jl --train "../../machine learning datasets/default/fisher_iris.csv" -c examples/configs/config_iris.toml  # override train path

Configuration

All options can be set via TOML config (preferred) or CLI flags. TOML overrides defaults; flags override TOML.

TOML structure

Parameter	TOML	Definition
train	train = "path.csv"	Training CSV (or single dataset to split)
test	test = "path.csv"	Test CSV (for testing.fixed)
testing	[testing.fixed] or [testing.split]	How to obtain test set
testing.fixed	[testing.fixed] + test = "..."	Use separate test file
testing.split	[testing.split] + split = [0.8, 0.2]	Split train ratio : test ratio
classifier	classifier = "dt"	dt, knn, or svm
parameters	[parameters] + k = 5	Classifier hyperparameters (e.g. k for KNN)
distance	distance = "euclidean"	Distance metric for KNN
goal.t	[goal] + t = 0.95	Accuracy threshold (0–1)
direction	[direction.forward] or [direction.backward]	Forward (additive) or backward (subtractive)
splits	splits = 1	Number of train/test splits
n	n = 10	Iterations per split
m	m = 5	Cases added/removed per iteration
sampling	[sampling.stratified] or [sampling.random]	Sampling method
seed	seed = 42	PRNG seed
output_dir	output_dir = "results"	Output directory

CLI flags (Python)

Flag	Description
-c, --config	TOML config file
--train	Training CSV
--test	Test CSV (for fixed)
--testing	fixed | split | cv
--split	Train,test ratio, e.g. 0.8,0.2
--classifier	dt | knn | svm
--k	K for KNN (default 3)
--distance	Metric for KNN
-t, --threshold	Accuracy threshold
--direction	forward | backward
--splits	Number of splits
-n, --iterations	Iterations per split
-m	Cases per iteration
--sampling	random | stratified
--seed	Random seed
-o, --output-dir	Output directory

Data format

CSVs must have a class column whose header matches class, label, or target (case-insensitive).
All other columns are features (column order may follow your benchmark file, e.g. fisher_iris.csv).

• fisher_iris.csv: class column
• mnist_train.csv, mnist_test.csv: label column

Classifiers

Code	Classifier
dt	Decision Tree
knn	K-Nearest Neighbors
svm	Support Vector Machine (RBF)
hb_vis	Hyperblock (VisCanvas-style)
hb_dv	Hyperblock (DV-style, interval-based)

Output

Results are written to {output_dir}/bap_{timestamp}/ (default results/bap_YYYYMMDD_HHMMSS/).

Exported case and hyperblock CSVs use the same tabular shape as the shared datasets (e.g. computing/machine learning datasets/default/fisher_iris.csv):

• One header class (lowercase), case label for data rows.
• One column per attribute (same names as the training CSV). No separate *_min / *_max columns and no extra ID column.
• split_N/converged_exp_{id}_seed{seed}.csv – converged training cases: class holds the dataset label (e.g. Setosa).
• split_N/converged_exp_{id}_seed{seed}_hyperblocks.csv – when using hb_vis or hb_dv: two rows per hyperblock. Each row has the same attribute columns; values are the box minimum (…__bottom) and maximum (…__top) corners. The class cell encodes label, HB id, and edge, e.g. Setosa__HB0__bottom / Setosa__HB0__top. Any __ in the dataset label is replaced by _ so the suffix pattern stays parseable.

Other output:

• config.txt – Settings used
• statistics.csv – Aggregate statistics (mean/min/max cases, convergence rate, etc.)

Every converged result CSV has a matching _hyperblocks.csv in the same directory when using hyperblock classifiers.

Algorithm (summary)

1. Set PRNG seed
2. For each split: load data (fixed test or split)
3. For each iteration: start with empty set (forward) or full set (backward)
4. While accuracy < threshold and cases remain: add/remove m cases via sampling
5. Record converged subsets and compute statistics

Bidirectional Processing Definition Notes

This repository originally exposed BAP behavior through rebuild.py; that interface has now been folded into the TOML/CLI model used by bap.py and julia/bap.jl.

Parameter mapping

• Legacy --data maps to train
• Legacy --test-data maps to test with testing.fixed
• Legacy train/test split flags map to testing.split with split = [train_ratio, test_ratio]
• Legacy --action {additive,subtractive} maps to direction {forward,backward}
• Legacy --iterations maps to n
• Legacy --threshold maps to goal.t

Formal algorithm input

Core BAP input fields are:

• train, test/testing, classifier, parameters, distance
• goal.t, direction, splits, n, m
• sampling, seed

Formal algorithm output

• Run configuration export (config.txt)
• Converged case-set CSV artifacts per successful iteration
• Aggregate statistics including convergence rate and sureness-related measures

Procedure (expanded)

1. Initialize PRNG with seed.
2. Repeat for each split:
   • Build train/test partitions (testing.fixed or testing.split).
3. Repeat n times:
   • Start with empty set (direction.forward) or full train set (direction.backward).
   • Train/test until threshold goal.t is met, adding/removing m cases each step.
   • Mark iteration as failed if no cases remain before meeting threshold.
   • Persist converged set and increment seed on success.