[Bug]: Multi-metric outcome constraints are silently reduced to a single metric #5309

What happened?

A linear outcome constraint over several metrics, such as "a + 2 * b <= 1.0", is accepted by Client.configure_optimization. When candidates are generated, though, all but one metric is dropped and the remaining one gets weight 1. There is no error or warning, so the optimizer works with a different constraint than the one the user specified.

Please provide a minimal, reproducible example of the unexpected behavior.

import ax.adapter.torch as torch_adapter
from ax.adapter.adapter_utils import extract_outcome_constraints
from ax.api.client import Client
from ax.api.configs import RangeParameterConfig

captured = []
def spy(outcome_constraints, outcomes):
    result = extract_outcome_constraints(outcome_constraints, outcomes)
    captured.append((outcomes, result))
    return result
torch_adapter.extract_outcome_constraints = spy

client = Client()
client.configure_experiment(
    parameters=[RangeParameterConfig(name="x", parameter_type="float", bounds=(0.0, 1.0))]
)
client.configure_optimization(objective="obj", outcome_constraints=["a + 2 * b <= 1.0"])

while not captured:  # run until the model-based node is used
    for idx, params in client.get_next_trials(max_trials=1).items():
        x = params["x"]
        client.complete_trial(trial_index=idx, raw_data={"obj": x, "a": x, "b": 1 - x})

outcomes, (A, b) = captured[-1]
print(outcomes, A)

Expected: one row of A with non-zero entries for both a and b, in the ratio 1 : 2.

Actual:
['a', 'b', 'obj'] [[0. 1. 0.]]
Only b is left, with weight 1.

Please paste any relevant traceback/logs produced by the example provided.

Ax Version

1.3.1

Python Version

3.12

Operating System

Linux Mint

(Optional) Describe any potential fixes you've considered to the issue outlined above.

Cause

extract_outcome_constraints (ax/adapter/adapter_utils.py) only applies all of a constraint's weights when the constraint is an instance of ScalarizedOutcomeConstraint. Otherwise it uses c.metric_signatures[0] with weight 1. ScalarizedOutcomeConstraint is now deprecated in favour of OutcomeConstraint(expression=...), and optimization_config_from_string creates a plain OutcomeConstraint for multi-metric expressions. Those constraints therefore always take the single-metric branch.

The same isinstance check in Derelativize._compute_sq_val (ax/adapter/transforms/derelativize.py) has the same effect on relative constraints. For example, "a + b >= 0.9 * baseline" is accepted, even though the docstring of configure_optimization says scalarized constraints cannot be relative. Its bound is then derived from only the first metric's status quo value.

Suggested fix

In both places, branch on len(c.metric_weights) > 1 instead of isinstance(c, ScalarizedOutcomeConstraint), and use c.metric_weights for the coefficients. Separately, reject relative constraints over several metrics when parsing, if they're not meant to be supported.

Pull Request

None

Code of Conduct

• I agree to follow Ax's Code of Conduct