haraldschilly · haraldschilly · May 28, 2026
diff --git a/TODO.md b/TODO.md
@@ -2,6 +2,63 @@
 
 ## Recent Improvements (continued)
 
+### NL-SHADE-LBC adaptive DE (CEC 2022 winner) — 2026-05-28
+- [x] **New `NLSHADE_LBC` heuristic** in
+      `panobbgo/heuristics/nl_shade_lbc.py`; closes the *NL-SHADE-LBC*
+      DE-family follow-up in `planning/SELF_IMPROVEMENT_LOOP.md`.
+      Direct subclass of
+      :class:`~panobbgo.heuristics.nl_shade_rsp.NLSHADE_RSP` (Stanovov,
+      Akhmedova & Semenkin, CEC 2022 winner) adding **Linear Bias
+      Change** in the success-history memory update: the F / CR
+      Lehmer-mean order ``p`` is linearly scheduled across budget
+      progress instead of fixed at 2.  Literature defaults
+      ``p_F: 3.5 → 1.5``, ``p_CR: 1.0 → 1.5``, spread ``m_lbc = 1.5``;
+      at ``p = 2, m = 1`` the formula recovers the standard L-SHADE
+      Lehmer mean.
+  - **Why it matters.** NL-SHADE-LBC won the CEC-2022 single-objective
+    bound-constrained competition and is the direct NL-SHADE-RSP
+    descendant — the literature frontier as of the most recent CEC
+    competition we can mirror.  Adds a fifth DE-family arm the bandit
+    can pick whichever wins on the current battery.
+- [x] **Catalog wiring** — `default_structural_catalog` gains
+      `(NLSHADE_LBC, {"NP_init": 30, "k_rank": 3.0})` as a fifteenth
+      `add_heuristic` candidate; `default_catalog` gains six
+      LBC-specific rules (`NLSHADE_LBC.NP_init` integer_add,
+      `NLSHADE_LBC.p_F_init` / `p_F_final` / `p_CR_init` / `p_CR_final`
+      float_uniform, `NLSHADE_LBC.m_lbc` float_uniform).
+- [x] **Deviations from the full CEC-2022 paper** (documented for
+      honesty): the adaptive binomial / exponential crossover blend
+      and the repetitive-generation bound-constraint handling are
+      *not* ported — the same async-pipeline limitations that motivated
+      omitting them from NL-SHADE-RSP apply here.  Both queued as
+      follow-ups.
+- [x] **CR-zero handling** — preserves the L-SHADE terminal sentinel
+      rule and filters strict zeros out of the LBC sum (because
+      ``s^(p−m)`` with ``p < m`` is undefined at ``s = 0``).
+- [x] **Backwards compatibility** — strictly safe.  NLSHADE_LBC is
+      opt-in: not added to any default battery, so existing composite
+      baselines stay byte-identical.  NL-SHADE-RSP / jSO / L-SHADE
+      base classes are untouched — only the LBC subclass overrides
+      `_update_memory`; verified by a regression test that
+      ``NLSHADE_RSP._update_memory`` still produces the standard
+      L-SHADE Lehmer mean output.
+- [x] **Tests** — `tests/test_heuristic_nl_shade_lbc.py` (30 tests):
+      construction validation, LBC schedule (endpoints, linear
+      midpoint, clipping, budget-unknown fallback), memory update
+      (anchor-bin skip, pointer mod (H-1), no-op on empty buffer,
+      [0,1] clamping, formula recovers Σ(w·F^3.5)/Σ(w·F^2.0) at
+      progress=0, p=2/m=1 recovers L-SHADE for both F and CR, CR=0
+      terminal sentinel, terminal-bin stays terminal, mixed-zero CR
+      filtered, uniform weights on zero-delta), pipeline (on_start,
+      smoke convergence, restart resets), inheritance safety
+      (NLSHADE_RSP unchanged), and registration.
+- [x] **Documentation updated** —
+      `planning/SELF_IMPROVEMENT_LOOP.md` (new §13 entry, follow-up
+      idea seeded), `doc/source/guide.rst`,
+      `doc/source/guide_benchmarking.rst`,
+      `doc/source/guide_architecture.rst`,
+      `doc/source/heuristics.rst`, and this TODO entry.
+
 ### NL-SHADE-RSP adaptive DE (CEC 2021 winner) — 2026-05-25
 - [x] **New `NLSHADE_RSP` heuristic** in
       `panobbgo/heuristics/nl_shade_rsp.py`; closes the *NL-SHADE-RSP /

diff --git a/doc/source/guide.rst b/doc/source/guide.rst
@@ -41,7 +41,7 @@ Quick Navigation
    * - **Interested in research?**
      - Explore :doc:`guide_research` for related work, theoretical properties, and future directions
    * - **Want to measure progress?**
-     - Read :doc:`guide_benchmarking` for the composite score, external baselines, parametrically randomised problems, the statistical acceptance rule (with paired and unpaired bootstrap sampling — paired is auto-selected for the rep-aligned randomized harness and shrinks the CI 3-10× over the historical independent-resample scheme), the autonomous self-improvement loop driver, the anti-cherry-pick guard, the hold-out validation set (single- and multi-seed, with bootstrap-CI aggregation across hold-out seeds), the adaptive Thompson-sampling mutation sampler with optional per-class structural bandit arms, the structural ``add_heuristic`` / ``drop_heuristic`` portfolio mutations, the categorical ``MutationRule`` kind (for discrete knobs like ``PSO.topology``, ``Sobol.scramble``, ``LSHADE.archive_factor``, ``LSHADE.F_schedule``, and ``NLSHADE_RSP.adaptive_archive``), the tri-topology PSO (``gbest`` / ``lbest`` / ``vonneumann`` — fully-connected, ring, and 4-connected 2-D toroidal grid per Kennedy & Mendes 2003 / Mendes 2004) candidate pool, the L-SHADE adaptive Differential Evolution heuristic (Tanabe-Fukunaga 2014) with two opt-in jSO refinements (the linearly-decreasing ``p_best`` schedule from iLSHADE / jSO Brest et al. 2016 / 2017 and the three-phase asymmetric F-cap from jSO Brest et al. 2017), the literature-faithful jSO heuristic itself (Brest, Maučec & Bošković 2017 — CEC-2017 winner — inheriting the L-SHADE F-cap machinery by construction), the NL-SHADE-RSP heuristic (Stanovov, Akhmedova & Semenkin 2021 — CEC-2021 winner — a jSO subclass adding non-linear population reduction, rank-based selective pressure, and a randomised adaptive archive), and the COBYQA derivative-free trust-region local optimizer (Ragonneau-Zhang 2023)
+     - Read :doc:`guide_benchmarking` for the composite score, external baselines, parametrically randomised problems, the statistical acceptance rule (with paired and unpaired bootstrap sampling — paired is auto-selected for the rep-aligned randomized harness and shrinks the CI 3-10× over the historical independent-resample scheme), the autonomous self-improvement loop driver, the anti-cherry-pick guard, the hold-out validation set (single- and multi-seed, with bootstrap-CI aggregation across hold-out seeds), the adaptive Thompson-sampling mutation sampler with optional per-class structural bandit arms, the structural ``add_heuristic`` / ``drop_heuristic`` portfolio mutations, the categorical ``MutationRule`` kind (for discrete knobs like ``PSO.topology``, ``Sobol.scramble``, ``LSHADE.archive_factor``, ``LSHADE.F_schedule``, and ``NLSHADE_RSP.adaptive_archive``), the tri-topology PSO (``gbest`` / ``lbest`` / ``vonneumann`` — fully-connected, ring, and 4-connected 2-D toroidal grid per Kennedy & Mendes 2003 / Mendes 2004) candidate pool, the L-SHADE adaptive Differential Evolution heuristic (Tanabe-Fukunaga 2014) with two opt-in jSO refinements (the linearly-decreasing ``p_best`` schedule from iLSHADE / jSO Brest et al. 2016 / 2017 and the three-phase asymmetric F-cap from jSO Brest et al. 2017), the literature-faithful jSO heuristic itself (Brest, Maučec & Bošković 2017 — CEC-2017 winner — inheriting the L-SHADE F-cap machinery by construction), the NL-SHADE-RSP heuristic (Stanovov, Akhmedova & Semenkin 2021 — CEC-2021 winner — a jSO subclass adding non-linear population reduction, rank-based selective pressure, and a randomised adaptive archive), the NL-SHADE-LBC heuristic (Stanovov, Akhmedova & Semenkin 2022 — CEC-2022 winner — a NL-SHADE-RSP subclass adding **Linear Bias Change** in the success-history memory update: the F / CR Lehmer-mean exponents are linearly scheduled across budget progress instead of fixed), and the COBYQA derivative-free trust-region local optimizer (Ragonneau-Zhang 2023)
 
 Guide Contents
 --------------

diff --git a/doc/source/guide_architecture.rst b/doc/source/guide_architecture.rst
@@ -298,8 +298,21 @@ Implemented Heuristics
   Inherits the jSO mutation / schedule / anchor-memory machinery and the asynchronous pipeline
   unchanged.  The full CEC-2021 adaptive-crossover blend and success-ratio archive-probability
   adaptation are intentionally not ported (the asynchronous generation model does not expose them
-  cleanly).  All four DE-family arms (DE / L-SHADE / jSO / NL-SHADE-RSP) ship in the structural
-  catalog so the bandit picks whichever wins on the current battery.
+  cleanly).
+
+- :class:`~panobbgo.heuristics.nl_shade_lbc.NLSHADE_LBC`: NL-SHADE-LBC refinement of NL-SHADE-RSP
+  (Stanovov, Akhmedova & Semenkin, CEC 2022 winner).  Direct subclass of
+  :class:`~panobbgo.heuristics.nl_shade_rsp.NLSHADE_RSP` that adds *Linear Bias Change* in the
+  success-history memory update: the F / CR Lehmer-mean order ``p`` is linearly scheduled across
+  budget progress (``p_F: 3.5 → 1.5``, ``p_CR: 1.0 → 1.5``) instead of fixed at ``2``; the spread
+  between numerator and denominator exponents is held constant at ``m_lbc = 1.5``.  At
+  ``p = 2, m = 1`` the formula recovers the standard L-SHADE weighted Lehmer mean — both regimes
+  are reachable from the default catalog so the bandit can flip between them.  Inherits NL-SHADE-RSP's
+  NLPSR / RSP / adaptive-archive machinery and the asynchronous pipeline unchanged.  The CEC-2022
+  adaptive crossover blend and repetitive-generation bound-constraint handling are intentionally not
+  ported (see the heuristic docstring).  All five DE-family arms (DE / L-SHADE / jSO / NL-SHADE-RSP /
+  NL-SHADE-LBC) ship in the structural catalog so the bandit picks whichever wins on the current
+  battery.
 
 - :class:`~panobbgo.heuristics.pso.PSO`: Asynchronous Particle Swarm Optimization with the canonical
   Clerc–Kennedy (2002) constriction-coefficient parameters (``w = χ ≈ 0.7298``, ``c1 = c2 ≈ 1.49618``).

diff --git a/doc/source/guide_benchmarking.rst b/doc/source/guide_benchmarking.rst
@@ -709,7 +709,7 @@ space with two new ops that change the *shape* of a
 * ``add_heuristic`` — append a heuristic from a curated pool
   (``Random``, ``Nearby``, ``NelderMead``, ``Center``,
   ``LatinHypercube``, ``Sobol``, ``Extremal``, ``PSO``, ``LSHADE``,
-  ``JSO``, ``NLSHADE_RSP``, ``COBYQA``) to a target strategy.  The pool ships *three*
+  ``JSO``, ``NLSHADE_RSP``, ``NLSHADE_LBC``, ``COBYQA``) to a target strategy.  The pool ships *three*
   PSO entries — the default fully-connected ``gbest`` topology
   (Kennedy-Eberhart 1995), the ring ``lbest`` topology with
   ``k_neighbors=2`` (Kennedy & Mendes 2002), and the 4-connected
@@ -734,7 +734,14 @@ space with two new ops that change the *shape* of a
   Non-Linear Population Size Reduction (``NP(r) = round((NP_min −
   NP_init)·r^(1−r) + NP_init)``), Rank-based Selective Pressure on the
   differential ``r1`` draw (``k_rank`` default ``3``), and a randomised
-  per-generation archive cap.  All four DE-family arms share the
+  per-generation archive cap.  NL-SHADE-LBC (Stanovov, Akhmedova &
+  Semenkin 2022 — CEC-2022 winner) refines NL-SHADE-RSP further with
+  **Linear Bias Change** in the success-history memory update: the F /
+  CR Lehmer-mean order ``p`` is linearly scheduled across budget
+  progress (``p_F: 3.5 → 1.5``, ``p_CR: 1.0 → 1.5``) instead of fixed
+  at ``2``; the numerator / denominator exponent spread ``m_lbc`` is
+  held constant at ``1.5`` (``m = 1`` and ``p = 2`` everywhere recovers
+  the standard L-SHADE Lehmer mean).  All five DE-family arms share the
   ``add_heuristic`` arm so the bandit picks whichever variant wins on
   the current battery.  COBYQA
   (Ragonneau-Zhang 2023) brings the modern Powell-family

diff --git a/doc/source/heuristics.rst b/doc/source/heuristics.rst
@@ -21,7 +21,8 @@ Point generation algorithms (heuristics):
 - **LSHADE**: Linear-population-reduction Success-History Adaptive DE (Tanabe & Fukunaga, CEC 2014 winner).  Adapts ``F`` and ``CR`` per-trial via per-bin Cauchy / Normal memories that update each generation by the weighted Lehmer mean of successful triples; uses the ``current-to-pbest/1`` mutation with an external archive of replaced parents.  The population shrinks linearly from ``NP_init`` (default 30) down to ``NP_min`` (default 4) over the strategy's evaluation budget — broad exploration early, focused exploitation late.  Out-of-bounds components are repaired by midpoint reflection.  Strictly stronger than the basic ``DE/rand/1/bin`` on multimodal benchmarks; opt-in via the structural mutation catalog.
 - **JSO**: jSO (Brest, Maučec & Bošković, CEC 2017 winner) — direct subclass of :class:`~panobbgo.heuristics.lshade.LSHADE` that adds three literature-best refinements to the success-history adaptive DE machinery: a **weighted current-to-pbest-w/1 mutation** (the pbest direction is re-weighted by a phase-dependent ``F_w`` factor — ``0.7·F`` early, ``0.8·F`` mid, ``1.2·F`` late), a **linear ``p_best`` schedule** that decreases from ``0.25`` to ``0.125`` over the budget, and **Cauchy-F clamping** that limits sampled ``F`` values to ``0.7`` while ``progress < 0.6``.  The history memory uses ``M_F = 0.3`` / ``M_CR = 0.8`` initial values and reserves the last bin (``H − 1``) as a frozen anchor at ``0.9 / 0.9`` that ``_update_memory`` never overwrites — a stable "moderately greedy" parameter source independent of live success history.  Inherits LSHADE's asynchronous pipeline (per-slot pending dict, generation-by-count book-keeping, archive trimming, LPSR shrinking, warm restart) unchanged.  Opt-in via the structural mutation catalog.
 - **NLSHADE_RSP**: NL-SHADE-RSP (Stanovov, Akhmedova & Semenkin, CEC 2021 winner) — direct subclass of :class:`~panobbgo.heuristics.jso.JSO` that adds three further refinements on top of the jSO machinery: **Non-Linear Population Size Reduction** (``NP(r) = round((NP_min − NP_init)·r^(1−r) + NP_init)``, which drops the population faster than L-SHADE's linear schedule in the early phase), **Rank-based Selective Pressure** (the differential ``r1`` index is drawn with probability proportional to a fitness rank weight ``k_rank·(n−i)/n + 1``, biasing the mutation toward better individuals; ``k_rank`` default ``3``), and a **randomised adaptive archive** (the archive cap is resampled per generation in ``[0, round(archive_factor·NP)]``).  Inherits jSO's weighted ``current-to-pbest-w/1`` mutation, linear ``p_best`` schedule, asymmetric F-cap, frozen anchor memory bin, and the asynchronous pipeline unchanged.  The full CEC-2021 adaptive-crossover blend and success-ratio archive-probability adaptation are intentionally not ported (see the heuristic docstring).  Opt-in via the structural mutation catalog.
-- **DifferentialEvolution / LSHADE / JSO / NLSHADE_RSP complementarity**: Panobbgo carries all four DE-family arms in the structural catalog so the bandit can pick whichever wins on the current battery — basic DE for byte-identical legacy reproduction, L-SHADE for the established CEC-2014 high-water mark, jSO for the CEC-2017 refinement, and NL-SHADE-RSP for the CEC-2021 refinement.  Each variant occupies a distinct exploration / exploitation regime and the literature shows real per-problem complementarity; on Panobbgo's small-budget battery NL-SHADE-RSP measures within noise of jSO (mean composite delta ≈ 0 across seeds, seed-dependent wins on both sides) — the CEC-DE refinements are large-budget specialists.
+- **NLSHADE_LBC**: NL-SHADE-LBC (Stanovov, Akhmedova & Semenkin, CEC 2022 winner) — direct subclass of :class:`~panobbgo.heuristics.nl_shade_rsp.NLSHADE_RSP` that adds **Linear Bias Change** in the success-history memory update.  The standard L-SHADE / jSO / NL-SHADE-RSP Lehmer mean uses fixed exponents (``s^2 / s^1``); NL-SHADE-LBC generalises the F / CR memory updates to ``Σ(w·s^p) / Σ(w·s^(p − m))`` with the *order* ``p`` linearly scheduled across budget progress — ``p_F(r) = (1 − r)·p_F_init + r·p_F_final`` (defaults ``3.5 → 1.5``) and ``p_CR(r) = (1 − r)·p_CR_init + r·p_CR_final`` (defaults ``1.0 → 1.5``).  The spread ``m`` is held constant at the literature default ``1.5``; at ``p = 2, m = 1`` the formula recovers the standard L-SHADE weighted Lehmer mean.  Inherits NL-SHADE-RSP's NLPSR, RSP, randomised adaptive archive, jSO's frozen anchor bin, and the asynchronous pipeline unchanged.  The full CEC-2022 adaptive crossover blend and repetitive-generation bound-constraint handling are intentionally not ported (see the heuristic docstring).  Opt-in via the structural mutation catalog.
+- **DifferentialEvolution / LSHADE / JSO / NLSHADE_RSP / NLSHADE_LBC complementarity**: Panobbgo carries all five DE-family arms in the structural catalog so the bandit can pick whichever wins on the current battery — basic DE for byte-identical legacy reproduction, L-SHADE for the established CEC-2014 high-water mark, jSO for the CEC-2017 refinement, NL-SHADE-RSP for the CEC-2021 refinement, and NL-SHADE-LBC for the CEC-2022 refinement.  Each variant occupies a distinct exploration / exploitation regime and the literature shows real per-problem complementarity; on Panobbgo's small-budget battery the CEC-DE refinements (RSP, LBC) measure within noise of jSO — they are large-budget specialists by construction.
 - **PSO**: Particle Swarm Optimization with the canonical Clerc–Kennedy (2002) constriction-coefficient parameters — a swarm of particles tracks personal and global bests with momentum, providing exploration dynamics distinct from CMA-ES (covariance) and DE (recombination); supports IPOP-style warm restarts via the Restart analyzer.  Three swarm topologies are available via the ``topology`` argument: the default fully-connected ``"gbest"`` (Kennedy-Eberhart 1995), the wrap-around ring ``"lbest"`` of width ``2·k_neighbors+1`` (Kennedy & Mendes 2002), and the 4-connected 2-D toroidal grid ``"vonneumann"`` (Kennedy & Mendes 2003; Mendes 2004) — three complementary information-diffusion regimes (instantaneous / one-hop linear / two-hop planar), with ``"vonneumann"`` widely cited as a stable middle ground that wins on a broader range of problem classes than either pure ring or pure star.  Optional ``w_end`` enables the Shi–Eberhart (1998) linearly-decreasing inertia schedule (off by default; constant ``w`` reproduces the prior behaviour).
 - **FeasibleSearch**: Actively searches for and repairs feasible solutions via line search towards feasibility.
 - **ConstraintGradient**: Uses estimated gradients of constraint violations to find feasible regions.

diff --git a/panobbgo/heuristics/__init__.py b/panobbgo/heuristics/__init__.py
@@ -44,6 +44,7 @@
 from .lshade import LSHADE
 from .jso import JSO
 from .nl_shade_rsp import NLSHADE_RSP
+from .nl_shade_lbc import NLSHADE_LBC
 from .repair import ConstraintRepair
 from .cma_es import CMAES
 from .pso import PSO
@@ -70,6 +71,7 @@
     "LSHADE",
     "JSO",
     "NLSHADE_RSP",
+    "NLSHADE_LBC",
     "ConstraintRepair",
     "CMAES",
     "PSO",