feat(nl_shade_lbc): NL-SHADE-LBC adaptive DE (CEC 2022 winner)

[haraldschilly]

Summary

Add NL-SHADE-LBC (Stanovov, Akhmedova & Semenkin, CEC 2022 winner) — the literature frontier of the SHADE / L-SHADE / jSO / NL-SHADE-RSP adaptive Differential Evolution lineage. It is a direct subclass of 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 (s²/s¹); NL-SHADE-LBC generalises this to Σ(w·sᵖ) / 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
p_CR(r) = (1 − r) · p_CR_init + r · p_CR_final   # defaults 1.0 → 1.5

with constant spread m_lbc = 1.5. Literature defaults from Stanovov et al. (2022) — verified against the MetaBox open-source reference implementation. At p = 2, m = 1 the formula recovers the standard L-SHADE Lehmer mean, so both regimes are reachable from the default catalog and the bandit can flip between them.

Why

Closes the NL-SHADE-LBC (CEC 2022 winner) DE-family follow-up in planning/SELF_IMPROVEMENT_LOOP.md. Adds a fifth DE-family arm to the structural catalog — alongside basic DE / L-SHADE (CEC 2014) / jSO (CEC 2017) / NL-SHADE-RSP (CEC 2021) — so the self-improvement loop's bandit can pick whichever wins on the current battery. Brings panobbgo to the most recent CEC bound-constrained competition winner.

Backwards compatibility

Strictly safe. NLSHADE_LBC is opt-in:

• Not added to any _make_quick_strategies / _make_standard_strategies / _make_full_strategies spec, so the composite baseline on every default battery is byte-identical and existing ledgers stay valid.
• NL-SHADE-RSP / jSO / L-SHADE base classes are untouched — only the LBC subclass overrides _update_memory; a regression test verifies that NLSHADE_RSP._update_memory still produces the standard L-SHADE Lehmer mean output.
• The structural catalog gains one extra add_heuristic candidate (avoid_duplicates=True).
• The six new kwarg rules fire only when a spec sets the matching kwarg explicitly.

Impact

Like the prior CEC-DE refinements (NL-SHADE-RSP), NL-SHADE-LBC is a large-budget specialist — at panobbgo's small composite-battery budgets (75–500 evals) the bias-change schedule barely warms up, so the quick-mode signal is expected within noise. The value of shipping it today is to give the self-improvement loop the literature-frontier arm the bandit can select once it has accumulated per-arm reward history.

Evidence form (per AGENTS.md "Agent-driven improve X PRs"): catalog-only addition; backwards-compatible (composite baseline byte-identical, existing ledgers stay valid); queued for nightly loop validation via the structural catalog.

Deviations from the full CEC-2022 paper

For transparency (the Panobbgo norm is literature-faithful ports), two mechanisms are intentionally not ported:

• The adaptive binomial / exponential crossover blend (the same async-pipeline limitation that motivated omitting it from NL-SHADE-RSP).
• The repetitive-generation bound-constraint handling (Panobbgo's asynchronous pipeline runs through strategy.constraint_handler and the L-SHADE midpoint-reflection repair instead).

Both are queued as follow-ups in planning/SELF_IMPROVEMENT_LOOP.md.

Test plan

• tests/test_heuristic_nl_shade_lbc.py (30 tests): construction validation (defaults, custom kwargs, subclass invariant spanning NLSHADE_RSP / JSO / LSHADE, NaN / inf / m_lbc≤0 rejection, inherited NLSHADE_RSP / jSO H≥2 / p_best ordering / k_rank rules); LBC schedule (endpoints, linear midpoint, clipping at progress > 1, fallback to p_init when budget unknown); memory update (no write to anchor bin H-1, pointer advances % (H-1), no-op on empty buffer, F memory clamped to [0,1], LBC formula at progress=0 matches Σ(w·F^3.5)/Σ(w·F^2.0), p=2/m=1 recovers L-SHADE for both F and CR, CR=0 plants terminal sentinel, terminal-bin stays terminal, mixed-zero CR values filtered, zero-delta successes fall back to uniform weights); pipeline (on_start emits NP_init, smoke convergence on a quadratic, restart resets archive and pending); inheritance safety (NLSHADE_RSP _update_memory still produces standard L-SHADE mean); registration (package re-export + __all__, structural catalog, six kwarg dials).
• tests/test_heuristic_nl_shade_rsp.py (34 tests), tests/test_heuristic_jso.py (36 tests), tests/test_heuristic_lshade.py (99 tests) — all pre-existing DE-family tests pass unchanged.
• tests/test_self_improve.py (180 tests) — structural catalog and kwarg rule additions don't break anything.
• Full pytest tests/ (1187 tests) — all pass.
• ruff format --check, ruff check, flake8 --select=E9,F63,F7,F82, pyright all clean.

Docs

heuristics.rst, guide_architecture.rst, guide_benchmarking.rst, guide.rst, TODO.md, and planning/SELF_IMPROVEMENT_LOOP.md (§13 entry + NL-SHADE-LBC follow-ups: categorical LBC regimes, per-CR / per-F sub-regime A/B, adaptive bias bounds from the success history).

https://claude.ai/code/session_01QubVM5JJdGDw2EhtZJFY8n

---

Generated by Claude Code