Cuts are still built from symbolic Jacobians (differentiate_mode ∈ {reverse_symbolic, sympy}, util.py:42-66), re-evaluated per iteration; PyNumero is used only for the Lagrangian regularization objectives (util.py:340-380, integers relaxed first). The 2020 "use AD instead of symbolic differentiation / talk to Michael Bynum" item remains open.
Scope:
- Assemble OA cut expressions from PyNumero numeric Jacobian rows at each trial point (
PyomoNLP.evaluate_jacobian()), keeping symbolic as fallback; benchmark expression-tree size vs evaluation cost on large instances (the historical motivation: models where symbolic cuts blow up).
- MindtPy already consumes grey-box AD (
cut_generation.py:183-215) — unify the two paths.
- Note
MAX_SYMBOLIC_DERIV_SIZE = 1000 heuristic in GDPopt loa.py does something similar (sympy below, reverse-numeric above) — share the policy.
Part of roadmap epic #2 (item C2).
Cuts are still built from symbolic Jacobians (
differentiate_mode ∈ {reverse_symbolic, sympy},util.py:42-66), re-evaluated per iteration; PyNumero is used only for the Lagrangian regularization objectives (util.py:340-380, integers relaxed first). The 2020 "use AD instead of symbolic differentiation / talk to Michael Bynum" item remains open.Scope:
PyomoNLP.evaluate_jacobian()), keeping symbolic as fallback; benchmark expression-tree size vs evaluation cost on large instances (the historical motivation: models where symbolic cuts blow up).cut_generation.py:183-215) — unify the two paths.MAX_SYMBOLIC_DERIV_SIZE = 1000heuristic in GDPoptloa.pydoes something similar (sympy below, reverse-numeric above) — share the policy.Part of roadmap epic #2 (item C2).