Scalarized Objective Optimizations with Ax
In some cases, you may want to optimize a linear combination of multiple metrics rather than a single metric. This is where scalarized objectives come into the picture. You can define an objective function that is a weighted sum of several metrics, allowing you to balance different aspects of performance in your optimization.
Scalarized objectives are useful when you have multiple metrics that you want to consider simultaneously in your optimization process. By assigning weights to each metric, you can control their relative importance in the overall objective function.
Using a scalarized objective is a good choice if you have a good idea of what the weights in the scalarization should be. If you're instead unclear about what the tradeoffs between the individual objectives are and want to run the optimization to understand those better, you should instead run a proper multi-objective optimization.
Setup
Before we begin you must instantiate the Client and configure it with your
experiment and metrics.
We will also assume you are already familiar with basic Ax usage.
client = Client()
client.configure_experiment(...)
client.configure_metrics(...)
Steps
- Configure an optimization with a scalarized objective
- Continue with iterating over trials and evaluating them
- Observe optimal parametrizations
1. Configure an optimization with a scalarized objective
We can leverage the Client's configure_optimization method to configure a
scalarized objective optimization. This method takes in an objective goal as a
string, and can be used to specify single-objective, scalarized-objective, and
multi-objective goals. For this recipe, we will use a scalarized-bjective goal:
client.configure_optimization(objectives="2 * objective1 + objective")
In this example, we are optimizing a linear combination of two objectives,
objective1 and objective2, and we value improvements to objective1 twice
as much as improvements in objective2.
By default, objectives are assumed to be maximized. If you want to minimize an
objective, you can prepend the objective with a -.
2. Continue with iterating over trials and evaluating them
Now that your experiment has been configured for a multi-objective optimization, you can simply continue with iterating over trials and evaluating them as you typically would.
# Getting just one trial in this example
trial_idx, parameters = client.get_next_trials(max_trials=1)().popitem()
client.complete_trial(...)
3. Observe optimal parametrizations
You can now observe the optimal parametrizations by calling
get_best_parameterization(). The function returns a list of tuples containing
the best parameters, their corresponding metric values, the most recent trial
that ran them, and the name of the best arm.
best_parameterization = client.get_best_parameterization()
for parameters, metrics, trial_index, arm_name in best_parameterization:
...
Learn more
Take a look at these other resources to continue your learning: