Decision Tree — Machine Learning/Classical Models

Decision Tree

Machine Learning / Classical Models

A fitted decision tree, evaluated at inference. The tree arrives as five flat columns – exactly the arrays scikit-learn's tree_ exposes – and the block walks it from the root: at each internal node it compares one feature against that node's threshold, going left when u[Feature] ≤ Threshold and right otherwise, until it reaches a leaf and outputs that leaf's value.

It is the workhorse of interpretable machine learning, and unlike almost everything else in this family it involves no transcendental and no arithmetic at all – only comparisons – so it is exact on every target, including hardware.

Ports

• u – the feature column, [d,1]. Every entry of Feature Index must name a row of it; d itself comes from whatever is connected, and is checked against the largest index used.
• Output – y, a [1,1] scalar carrying the value of the leaf reached. For a regression tree that is the predicted value; for a classifier, the predicted class label as you stored it in Leaf Value.

Parameters

All five are columns of the same height N, one row per node, with the root at row 0. This is the layout scikit-learn already gives you: tree_.feature, tree_.threshold, tree_.children_left, tree_.children_right and tree_.value.

• Feature Index – which entry of u the node tests, zero-based. Ignored at leaves, where scikit-learn stores −2.
• Threshold – the value it is tested against. Ignored at leaves.
• Left Child – the row taken when u[Feature] ≤ Threshold. A negative entry marks the node as a leaf, which is scikit-learn's own convention (children_left is −1 there).
• Right Child – the row taken otherwise. Ignored at leaves.
• Leaf Value – the output produced when the walk stops at this node. Ignored at internal nodes.
• Sampling Time (s) – zero or less inherits the solver's rate; a positive value runs the block at that period.

Children must have a higher row number than their parent. Every fitted tree exporter emits them that way, and requiring it means a malformed table is reported at the start of the run rather than sending the code generator into an endless walk.

Code export

All ten targets: Python, MATLAB, Java, Rust, C, C++, VHDL, Verilog, SystemVerilog and PLC Structured Text. Every target is exact rather than merely within tolerance, and the three HDL targets are genuinely synthesizable: the block only compares and assigns.

The tree is emitted as a nested if cascade rather than as a loop over the five arrays. That is what keeps it synthesizable – a literal walk would need a data-dependent index into five arrays on every iteration, which in hardware is a multiplexer over every node, once per level. As a cascade each node appears exactly once and no index arithmetic exists at all, so the emitted code is proportional to the number of nodes rather than to nodes × depth.

Simulink bridge

None. The reference would be the Statistics and Machine Learning Toolbox and its fitctree / predict blocks, which is not installed on this machine – so a bridge could not be run against a parity testbench even if one were written. Those blocks also take a fitted model object from a workspace variable, which no parameter mapping can carry: that is the same obstacle that makes this whole family Support::None, and it is the reason the tree is pasted here as plain columns instead. The bridge reports the block rather than dropping it silently, and it has no parity testbench, which is the documented consequence of Support::None rather than a gap. Code export verification still covers it across all ten languages.

Notes

• Algebraic and stateless: the output depends only on the current sample.
• Piecewise constant. The output changes only when a sample crosses a split, so it is a staircase, not a curve – and its derivative is zero almost everywhere. That is worth remembering before putting one inside a feedback loop.
• No state space, deliberately, for exactly that reason: no A/B/C/D is true of a piecewise-constant map. Model reduction reports the block as unmergeable, which is the honest answer.
• The comparison is ≤ going left, matching scikit-learn. A sample landing exactly on a threshold therefore goes left, and in fixed point that is not a rare case – an HDL port delivers Q16.16, so a threshold that is exactly representable is landed on rather than approached.
• Random Forest, Gradient Boosted Trees and Isolation Forest Score are this same walk with a different accumulator over a stacked table.