Add Constraint primitive; replace set_outcome_constraints string DSL

[499602D2]

Replaces the set_outcome_constraints string DSL with a proper Constraint primitive modeled as a peer of Objective. Constraint metrics register as tracking metrics on the Ax experiment before the OutcomeConstraint references them, so their values flow through raw_data alongside objectives and Ax never silently marks a trial failed for "missing required metric". Bounds are absolute; Ax's OutcomeConstraint defaults to relative, which would silently change the meaning of every gte/lte supplied.

Addresses #24. The pattern Bulut tried in the comment thread (register a metric as both an objective and a constraint via the string DSL) crashed Ax with "Cannot constrain on objective metric", and the alternative of referencing an unregistered metric name silently marked every trial failed. Constraint is the peer-of-Objective registration asked for in the thread.

Self-bounds on Objective (Objective(..., lte=X)) as sketched in #24 are a follow-up.

Migration

# Before: string DSL. Only worked if the constraint name was also a
# registered objective, which then triggered
# "Cannot constrain on objective metric".
session.backend.set_objectives([objective, Objective(name="pressure", ...)])
session.backend.set_outcome_constraints(["pressure >= 45.0"])

# After
pressure = Constraint(
    name="pressure", objective_function=read_pressure, gte=45.0,
)
session.backend.set_objectives([objective])
session.backend.set_constraints([pressure])

Validation

Constraint.__init__ rejects at construction:

• No bound (either gte or lte must be supplied)
• Inverted bounds (gte > lte)
• Single-point feasible range (gte == lte): meaningless for a soft outcome constraint on a continuous GP surrogate; use gte=X-eps, lte=X+eps for a tight tolerance
• Non-finite bounds (NaN, ±inf): use None for "no bound on this side"

Backend.set_constraints rejects at registration:

• Anything that is not a Constraint (an Objective smuggled in used to crash later with AttributeError; now rejects up front with a clear TypeError)

Tests

Unit tests cover every validation branch. Three canaries drive real Ax end-to-end:

• TestConstraintConvergenceCanary: BO on a 2D quadratic with x1 >= 2; asserts best feasible point is near the constrained optimum.
• TestMOOWithConstraint: two Objectives in Pareto MOO mode plus one Constraint. Asserts BO explores both ends of the feasible Pareto front.
• TestSessionLoopWithConstraint: drives Session.local_optimization end-to-end with a real AxBackend and a Constraint, staging cases in the archival dir between cycles. Catches orchestration bugs in the Case metadata to backend.tell handoff.

[499602D2]

@blttkgl this is the top PR, built on top of #25 and #26.

[blttkgl]

Hey @499602D2 ,

Things seems to work on my end. I pushed two commits, one is making changes in session.py to print/log the scalarizedObjectives and constraints, the other one is changing the aerofoil tutorial to use a lift constraint. If you could test it on your end I think we can merge this PR alongside #26 .

[blttkgl]

Having said that I do get a few warnings during the tutorial run, pasting them below. Any thoughts?

/home/blttkgl/flowboost/.venv/lib/python3.12/site-packages/gpytorch/likelihoods/noise_models.py:150: NumericalWarning: Very small noise values detected. This will likely lead to numerical instabilities. Rounding small noise values up to 1e-06.
  warnings.warn(
/home/blttkgl/flowboost/.venv/lib/python3.12/site-packages/botorch/optim/optimize.py:796: RuntimeWarning: Optimization failed in `gen_candidates_scipy` with the following warning(s):
[OptimizationWarning('Optimization failed within `scipy.optimize.minimize` with status 2 and message ABNORMAL: .'), OptimizationWarning('Optimization failed within `scipy.optimize.minimize` with status 2 and message ABNORMAL: .'), OptimizationWarning('Optimization failed within `scipy.optimize.minimize` with status 2 and message ABNORMAL: .'), OptimizationWarning('Optimization failed within `scipy.optimize.minimize` with status 2 and message ABNORMAL: .')]
Trying again with a new set of initial conditions.
  return _optimize_acqf_batch(opt_inputs=opt_inputs)
/home/blttkgl/flowboost/.venv/lib/python3.12/site-packages/botorch/optim/optimize.py:796: RuntimeWarning: Optimization failed on the second try, after generating a new set of initial conditions.
  return _optimize_acqf_batch(opt_inputs=opt_inputs)

[blttkgl]

I wonder if we should also add a random seed in the aerofoil tutorial, for better reproducibility? I had a test where I reached the best design by design 12, and current one is still stuck at a local maxima around design 20.. It just seems that the model is very big on exploitation, but exploration is secondary. Do you think this is due to non-deterministic seed values, or is there a way on Ax backend to prioritize exploration over exploitation, similar to e.g. GpyOPT?

The best design is somewhere along angleOfAttack ~10, but we are kinda stuck around 17.