You are implementing hyperparameter tuning by using Bayesian sampling for an Azure ML Python SDK v2-based model training from a notebook. The notebook is in an Azure Machine Learning workspace. The no

Setting max_concurrent_trials to 20 maximizes parallelism but defeats Bayesian optimization because nearly all trials start before any results are available, effectively reducing the search to random sampling.

Tightening the Bandit policy slack_factor from 0.2 to 0.1 terminates more underperforming trials earlier but does not change how Bayesian sampling selects new configurations or improve its convergence quality.

Setting slack_factor to 0.9 makes the Bandit termination policy extremely lenient, wasting compute on poor-performing trials and reducing overall tuning efficiency without benefiting Bayesian convergence.

Bayesian sampling works by using outcomes from previously completed trials to select the next hyperparameter configuration, so running fewer trials concurrently gives the algorithm more information per selection decision. With max_concurrent_trials set to 4 instead of 10, a greater proportion of trials complete before new ones are queued, which increases the quality of each selection and accelerates convergence. This is the recommended practice - keeping concurrency low relative to the total trial count maximizes the benefit of Bayesian optimization.