Decision Automation
Enhancing Model Performance through Incremental Development
Summary
A repetitive decision can often be made by a machine more consistently, effectively, quickly, and cheaply than by a human.
Furthermore, when a decision is automated it is much simpler to safely trial adjustments to a decision algorithm and obtain accurate results.
Example automated decision:
1. Risk estimation
2. Routing and QC level for bespoke manufacturing
3. Remediation level for customer complaint
.png)
Process
Overview
Decision automation typically starts with a training data set. This training set, including the input features and target selection, can be incrementally developed over time. Most of the improvements in model performance will likely come from developing this training set.
Model Creation
The actual process of creating the machine learning model can be quite quick, though it may be desirable to slow the process down in order to create a model whose sensitivity to different inputs is more directly explainable.
Experiment
When experimenting with adjustment to decision-making models, it is helpful if:
• New models are gradually deployed (blue-green paradigm) to enable safety monitoring
• New models can be safely deployed without the help of SWE