Incremental Classification Naive Bayes Fit — Machine Learning/Incremental Learning

Machine_Learning/Incremental_Learning/Incremental_Classification_Naive_Bayes_Fit · 2 input / 1 output port(s) at insert · exports to Python, MATLAB, Java, Rust, C, C++

Description#

The block's own DESCRIPTION_HTML, rendered verbatim — the same text the config dialog's info panel and the library navigator show. Fix a wrong sentence in the block's .cpp (R-D9), never here.

Code facts#

Ports#

Ports the constructor creates. A block whose port list changes with its configuration adds or removes ports at load time; the count above is the one a freshly inserted block has.

Configuration variables#

Every block also carries Sampling Time (s) from ICoreBlockSolverEnvironment: zero or less inherits the solver's rate, a positive value runs the block at that period.

Simulink bridge#

Caveat (shown to the user): Simulink's IncrementalClassificationNaiveBayes Fit takes its learner as InitialLearner, the name of an incrementalClassificationNaiveBayes object in the MATLAB workspace, which has no public constructor a generated script could write; the block is reported rather than dropped

Catalog contract: src/ICoreBlocks/ICoreCoder/ICoreCommandSystem/SimulinkBridge/ICoreSimulinkBlockCatalog.h

Description vs code#

The lists agree. check_block_descriptions.py finds no disagreement between the description's Ports, Parameters, Code export and Simulink bridge lists and the code's.

The verdict above is tools/docs/check_block_descriptions.py (P7.1), which compares LISTS. It cannot read a sentence: "stateless" on a block with a state, an initial-value semantic the recursion does not implement, a "not synthesizable" caveat the HDL banner contradicts. That is the agent audit (P7.3) on BLOCK_DESCRIPTION_AUDIT.md, and this tool's green is not a substitute for one.

File banner (developer view)#

The top comment of the block's .cpp — the maths, the realization and the export strategy, addressed to whoever changes it. It must not contradict the description above (P7.5).

IncrementalClassificationNaiveBayes Fit -- an online Gaussian naive Bayes, its model on a bus statsIncremental/Classification/NaiveBayes/IncrementalClassificationNaiveBayes Fit, MEASURED on R2026a 2026-10-02 (BLOCKS_TO_ADD_TOOLBOXES.md statsIncremental, on FEATURES_TO_ADD.md BF1). Simulink's block takes its learner as InitialLearner, the name of an incrementalClassificationNaiveBayes object in the MATLAB workspace; here the two options the block reads are configs, "Class Names" and "Metrics Warmup Period", with MATLAB's default warm-up (1000). The predictor count is x's width.

WHAT EACH STEP DOES, all of it measured against the Simulink block and the object's fit():

• IsWarm is decided from the observations BEFORE this step: true once Metrics Warmup Period

of them have arrived -- every step counts, a NaN or unknown label included -- so it flips one step after the object's fit() would set it (warm-up 3: step 4, in three runs).

• A label that is one of Class Names is counted for MajorityClass (1-based, the FIRST of a

tie, the current label included); a NaN or a label not in Class Names is ignored.

• The observation then updates its class: the count, and each predictor's running mean and

sum of squared deviations (Welford's update). The published standard deviation divides by n -- the POPULATION one, measured: two observations -0.8968 and 0.6531 give 0.7749 -- and is 0 for a class seen once; mean and deviation are 0 for a class not yet seen.

• Prior is each class's share of the fitted observations, and 1/K each before the first.
• CanPredict is true from the first step, as R2026a's is.

⚠ THREE DELIBERATE DIVERGENCES FROM THE SIMULINK BLOCK, all measured and all stated in the description:

1. R2026a's block NEVER FITS THE LAST CLASS in ClassNames. Its observations are counted for

MajorityClass and then dropped: with ClassNames [0 1] the Prior stays [1 0] for the whole run, and with [10 20 30 40] class 40 keeps a zero prior and zero statistics, where the object's own fit() learns every class. This block learns every class, as fit() does.

1. A NaN predictor reaches R2026a's running sums and turns that class's mean and deviation

into NaN for the rest of the run. Here an observation with a non-finite predictor is not fitted (its label is still counted for MajorityClass).

1. R2026a's block fits an observation whose label is NOT a class -- a NaN, a 7 -- as the

PREVIOUS valid label: labels 2 3 NaN 2 3 7 fitted x(3) and x(6) as class 3 and counted them for MajorityClass (with interpolation off, so it is the block, not its source), and two leading NaNs fitted nothing. Here such an observation fits nothing and counts for nothing but the warm-up, which the learner's fit() -- which refuses an unknown label -- supports better than a silent relabelling.

The bus is R2026a's with DistrNormal [K P 2] split into DistrMean and DistrStd [K,P] (ICoreIncrementalNaiveBayesSupport.h says why), and Prior a column [K,1] where Simulink's is 1-D [K].

Sample results#

This block carries a ICoreBus signal, whose value is text rather than a number and is not something a plot has an axis for. The samples are in the table below, exactly as the run recorded them.

Every 4th of 60 samples, from the table stimulus.

The same rig also ran:

Plotted: table — Repeating Sequence Stair: [-2 -1 -0.5 0 0.5 1 2 3], one entry per sample

Category static · sample time 0.1 · 60 steps · commit ab8631aa8 · produced by docsSample --out <folder> --blocks Incremental_Classification_Naive_Bayes_Fit Incremental_Classification_Naive_Bayes_Predict --steps 60 · data docs/generated/samples/Machine_Learning__Incremental_Learning__Incremental_Classification_Naive_Bayes_Fit.json · the SVG is generated from those numbers by tools/docs/plot_svg.py, so it is a run and not a drawing (R-D10).