[Bug]: get_next_trials violates parameter constraint

[kurt-essential]

What happened?

I set a constraint that four parameters must sum to <= 1 and the first trial generated sets all parameters to 0.5, violating this constraint. (on ax-platform v1.0.0)

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

Code:

from ax.api.client import Client
from ax.api.configs import RangeParameterConfig

parameter_names = [f'x_{i}' for i in range(4)]
parameters = [
    RangeParameterConfig(
            name=x,  
            bounds=(0.0, 1.0),
            parameter_type="float",
        ) for x in parameter_names]

sum_constraint = " + ".join(parameter_names) + " <= 1.0"
client = Client()
client.configure_experiment(
    parameters=parameters,
    parameter_constraints=[sum_constraint],
    )
client.configure_optimization(objective="objective")
print(sum_constraint)
print(client._experiment.search_space.parameter_constraints)
trials = client.get_next_trials(max_trials=3)
for trial_index, trial_parameters in trials.items():
    print(trial_index, trial_parameters)
    print(client._experiment.search_space.parameter_constraints[0].check(trial_parameters))

Output:

x_0 + x_1 + x_2 + x_3 <= 1.0
[ParameterConstraint(1.0*x_0 + 1.0*x_1 + 1.0*x_2 + 1.0*x_3 <= 1.0)]
0 {'x_0': 0.5, 'x_1': 0.5, 'x_2': 0.5, 'x_3': 0.5}
False
1 {'x_0': 0.07523129135370255, 'x_1': 0.07053725980222225, 'x_2': 0.32778964564204216, 'x_3': 0.3629995724186301}
True
2 {'x_0': 0.12477120012044907, 'x_1': 0.13883846905082464, 'x_2': 0.11183641757816076, 'x_3': 0.012117702513933182}
True

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

Ax Version

1.0.0

Python Version

3.10.12

Operating System

NAME="Ubuntu" VERSION="22.04.2 LTS (Jammy Jellyfish)"

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

No response

Pull Request

Yes

Code of Conduct

• I agree to follow Ax's Code of Conduct

[Balandat]

Ww recently made a change in the default strategy to evaluate the center of the search space in the first trial since we've found this to be quite helpful both on benchmarks as well as internal use cases.

Of course this is a very bad strategy if the center of the search space (when only looking at bounds) is infeasible. Thanks for flagging, this is something we'll fix.

[kurt-essential]

Thanks! Btw it would be nice if there was an easy way to turn off this option - specifically in cases where the user has prior trial data they want to start from, the center point may be less informative than for the cold start and user might not want it :)

[saitcakmak]

Hi @kurt-essential. You can configure / disable center point and initialization more broadly following this recipe: https://ax.dev/docs/next/recipes/influence-gs-choice#1-configure-a-generation-strategy-via-configure_generation_strategy

[kurt-essential]

Thanks @saitcakmak ! Would consider linking that in the "attach existing data" tutorial as it seems like a common scenario where one might want to do that