Roadmap

This document tracks the direction of cmenpy: what is in the next release, what is planned, and what is deliberately not going to be built here.

Status legend:

Status	Meaning
done	shipped in a release
in progress	under development on the current milestone
planned	agreed scope, not started
backlog	desirable, not scheduled

Design principles

• Vector first. The hot paths stay flat NumPy/Numba arrays; object wrappers are an interface, never the storage.
• Non-monolithic. No plotting, no forced logging, no heavyweight dependencies. Persistence and reporting are opt-in adapters (history=..., terminator=...) that never run unless asked for.
• Two declaration styles, one contract. @cm.declare and @cm.template keep producing the same solve(...) surface; new features extend both or neither, never only one.
• Backward compatibility inside a channel. Additive changes ship freely; renames keep a deprecated alias for one release (see cm.jit → cm.precompile).

2026b2 — the next release

One active milestone that bundles the work of feature request #23 (split into #33 and #34) with the run-control and analysis features requested alongside it.

Issue	Deliverable	Status
#33	Composable terminators (cmenpy.terminator)	planned
#34	Decoders and Space (cmenpy.decoder, cmenpy.space)	planned
#35	Run history, RAM + disk (cmenpy.history)	planned
#36	Multiobjective Pareto core	planned
#37	Per-dimension bounds (Bounds.lows/ups/sample/clip)	planned
#38	Changelog restructure + this roadmap	in progress
#39	Test functions: examples, docs and tests use opfunu-core / enoppy-core	in progress

Terminators (#33)

Stopping criteria become composable objects instead of hardcoded loop exits:

terminator = (
    cm.MaxEvaluation(nfe=50_000) & cm.MaxEpoch(limit=10_000)
) | cm.MaxPatience(ms=60_000)

model.solve(f, bounds, epochs=10_000, pop_size=40, terminator=terminator)

Python cannot overload && / ||, so & and | are the supported spelling. MaxPatience(ms=...) stops after N milliseconds without improvement and MaxTime(seconds=...) is the hard wall-clock cap. cm.Terminator(nfe=...) is kept as an alias of MaxEvaluation. Termination reasons surface in model.report.

Decoders and Space (#34)

The decision space stays a flat continuous vector, but each slice of it can now be interpreted as a different domain type — integers, strings, categories, permutations, binaries — through BaseDecoder subclasses.

space = cm.Space[
    cm.floats(low=0.0, up=1.0),
    cm.integers(low=0, up=10),
    cm.strings("A", "B", "C"),
]

# sugar form (a call, since Space[float=...] is not valid Python syntax)
space = cm.Space(float=(0.0, 1.0), string=("a", "b", "c"))

solve(space=...) is additive; bounds= keeps working, and ctx.space exposes the decoder pipeline that runs immediately before fitness evaluation.

Run history (#35)

Per-epoch records become storable in RAM or on disk, opt-in per run:

model.solve(..., history=True)             # in-memory summary
model.solve(..., history="run.jsonl")      # streamed summary
model.solve(..., history=cm.history.NPZ("run.npz", snapshots=True))

debug=True keeps its current full in-RAM behavior. A versioned reader (cm.history.load) returns the records without imposing pandas.

Multiobjective core (#36)

Vector-valued objectives end-to-end: per-objective sense, fitness stored as (pop, m) in the buffer, dominance/non-dominated sort/crowding in cmenpy.multiobjective, Pareto fronts on the population (pop.front, pop.crowding) and an NSGA-II reference example. Scalar runs keep their exact shapes and performance.

Per-dimension bounds (#37)

Bounds.low / Bounds.up expose only the first dimension, so the common rng.uniform(bounds.low, bounds.up, ...) idiom silently ignores unequal intervals. Bounds.lows, Bounds.ups, Bounds.sample() and Bounds.clip() fix this without breaking the existing accessors; examples and docs migrate to the new helpers.

Test functions (#39)

The engine ships without a benchmark suite — reference test functions come from the LTSIM companions opfunu-core (500+ numerical functions: classical name-based plus every CEC competition) and enoppy-core (50+ constrained engineering problems). Every example solves an opfunu function, example/engineering_demo.py shows enoppy, and tests/test_benchmarks.py keeps both integrations honest. See the Test functions guide.

Backlog (not scheduled)

• Checkpoint / resume: serialize buffer, mask, RNG state and epoch so a run can continue from disk.
• MOO quality indicators: hypervolume, IGD, front spread.
• Constraint handling (penalty / feasibility rules) on top of decoders.
• Batched and parallel target evaluation.
• A reproducible benchmark suite (CEC-style functions).
• Polished UserInterrupt signal handling.

Non-goals

cmenpy will stay a design and prototyping engine. We will not add: plotting or visualization, rigid logging frameworks, a model zoo of named algorithms as a core dependency, or non-GPU/non-NumPy numeric backends. Those belong in companion packages.

Proposing work

Open an issue first — see CONTRIBUTING.md. Ideas that are adopted land in the 2026b2 table above or in the backlog, and the changelog's Unreleased section is where they become visible.