Introduction

An R package for discriminative parameter learning of Bayesian Networks via Differential Evolution

dplbnDE trains Bayesian Network parameters using Differential Evolution (DE) to maximize Conditional Log-Likelihood (CLL) — a discriminative objective that directly minimizes classification error, rather than the standard log-likelihood used by generative learning.

Why discriminative learning?

Standard maximum likelihood estimation (MLE) for Bayesian Networks optimizes the joint probability of the data. This is the right objective for generative models, but not for classifiers. Any Bayesian Network structure encodes conditional independence assumptions that may not hold exactly in the data — this mismatch produces classification error. Discriminative learning directly targets CLL, which measures how well the model predicts the class given the features, minimizing that error.

Key capabilities

• 10 DE variants — classic, adaptive, and hybrid memetic algorithms
• Multiple BN structures — Naive Bayes, Tree Augmented Naive Bayes (TAN), Hill-Climbing, or custom topology
• Adaptive parameter control — algorithms like JADE and L-SHADE automatically tune mutation factor F and crossover rate CR
• Hybrid local search — SHADE-ILS, SHADE-MTS, MA-SW-Chains, and MOS combine global DE with Solis-Wets local refinement
• bnclassify integration — returned networks are bnc_bn objects compatible with predict(), cLogLik(), and other bnclassify functions

Algorithm overview

Algorithm	Category	Adaptation	Population Reduction
DEbest	Classic	Fixed F, CR	No
DErand	Classic	Fixed F, CR	No
JADE	Adaptive	Cauchy/Normal	No
L-SHADE	Adaptive	History-based	Linear (LPSR)
jSO	Adaptive	Phase-based	Linear (LPSR)
NL-SHADE-RSP	Adaptive	History-based	Non-linear
SHADE-ILS	Hybrid	History-based	No
SHADE-MTS	Hybrid	History-based	No
MA-SW-Chains	Hybrid	Fixed F, CR	No
MOS	Hybrid	History-based	No

Installation

Install dplbnDE from CRAN or GitHub

Quick Start

Train your first classifier in minutes