Set Stopping Conditions

This guide shows how to set stopping conditions for an optimization run for the Ax optimizer backend. These conditions can be a variety of methods such as meeting a certain toleration, reaching a maximum number of iterations, or exceeding a time limit

Defining Stopping Conditions

Ax has built-in stopping conditions, found here.

However, if you need a more customizable stopping condition, you can configure one following this format.

from ax.global_stopping.strategies.base import BaseGlobalStoppingStrategy
from ax.core.experiment import Experiment

class CustomStoppingStrategy(BaseGlobalStoppingStrategy):
    def __init__(
        self,
        min_trials: int = 1,
        inactive_when_pending_trials: bool = True,
        # Add any additional parameters you want to customize here
    ) -> None:
        super().__init__(
            min_trials=min_trials,
            inactive_when_pending_trials=inactive_when_pending_trials
        )
        # Initialize your custom parameters here

    def _should_stop_optimization(self, experiment: Experiment, **kwargs) -> tuple[bool, str]:
        # Implement your custom stopping logic here
        if your_condition_is_met:
            return True, "stopping criteria met"
        return False, "stopping criteria not met"

Adding Stopping Conditions to the Agent

The Agent can then be configured with an additional parameter, stopping_strategy, to use the custom stopping condition.

To run until the stopping condition is met, pass iterations=None to blop.ax.Agent.optimize(). Only use an unbounded run when the agent has a stopping strategy configured; otherwise the optimization has no termination condition.

agent = Agent(
    ...,
    stopping_strategy=CustomStoppingStrategy(),
)
RE(agent.optimize(iterations=None))

A finite iterations value remains a maximum: the stopping strategy may end the run before that limit is reached.