Explain why early stopping isn't implemented

[davystrong]

Hello,
I was looking at the documentation for BayesianParameterSampling for HyperDrive and there is a note saying that it doesn't support early stopping. It would be useful if you could explain why this is the case (I don't think there is any inherent limitation in the algorithm, but I could be wrong) or whether this is planned.

---

Document Details

⚠ Do not edit this section. It is required for docs.microsoft.com ➟ GitHub issue linking.

• ID: e06f0efe-069e-b703-eb92-6e5832f0a0b2
• Version Independent ID: 58afae14-065b-f63d-c679-8def0afb54fd
• Content: azureml.train.hyperdrive.BayesianParameterSampling class - Azure Machine Learning Python
• Content Source: AzureML-Docset/stable/docs-ref-autogen/azureml-train-core/azureml.train.hyperdrive.BayesianParameterSampling.yml
• Service: machine-learning
• Sub-service: core
• GitHub Login: @DebFro
• Microsoft Alias: debfro

[nagau]

Bayesian would need to look at the scores for already tried trials, to be able to recommend new models looking at priors.
In case hyperdrive early terminates a few trials, it won't know what would have been the score for those trials at end of training. Hence Bayesian won't be able to use those scores for recommendation. Because of this reason, For now, hyperdrive disallows early termination in bayesian.