When working with Rate projects that include Optim blocks, several cloning options are available to give you full control over how incorporated constraints behave. This flexibility is important because constraints fall into two main categories:
Constraints independent of dataset size, including the local constraints for minimum and maximum premium as well as proportion-based constraints such as Conversion Rate, Expected Loss Ratio, and Combined Ratio.
These constraints remain unchanged when cloning a project.
Constraints dependent on dataset size, known as extensive constraints, such as Premium Offered, Premium Written, and Margin Written.
For these, different cloning behaviors are available and described below.
When you clone a Rate project that contains at least one Optim block with a tagged result, the system will prompt you to choose among the available cloning options outlined below.
In the left hand side, you can navigate through all the optimization grid searches of the project, as well as the tagged results coming from them.
Note that, no matter which option is chosen for the cloning, it will apply to all the Optim tagged results.
Each of these options will have a different effect in the way the optimization is performed in the cloned Rate project:
Same: this option leaves the constraints unchanged, so the exact same values defined in the original project are applied to the new one. While it is the most straightforward approach, it may lead to unintended outcomes if the dataset used in the cloned project differs significantly from the original, as the constraints might no longer be appropriate or meaningful.
Rescale: in this case, the constraints will be recalculated to consider the size of the new database. A linear transformation is performed based on the number of lines of the databases. Let’s say we have a constraint on GWP set at 200M euros in the original project, which is built on a dataset containing 2M rows. If the project is cloned on another dataset with 3M rows, then the constraint will be changed to 300M euros. In all cases, you may see what the new values for the constraints will be by selecting the corresponding Optim block variant, both before cloning (using the list of grid searches in the left of the image above) and after (in the side bar of the block).
Score: when this option is selected, the constraints themselves are not directly reused. Instead, the Lagrange multiplier (λ) obtained from the original optimization problem is preserved and applied in the cloned project. In the original formulation, the optimization seeks to maximize an objective function f(x) subject to one or several constraints g(x). The Lagrange multiplier expresses the trade-off between the objective and each constraint according to the equation below:
The multiplier λ quantifies how strongly the optimization reacts to the constraint. Mathematically, it represents the rate of change in the optimal value of f(x) when g(x) is relaxed by one unit. In other words, λ measures how much the objective could improve if the constraint were slightly loosened.
The key advantage of this approach is consistency: the same optimization rule (through λ) is applied uniformly to every observation in the dataset. As a result, if the same policy appears in both the original and the cloned datasets, it will receive the same optimized price. This behavior differs from the previous cloning options, where constraints themselves are copied and a new optimization is solved on the new dataset, which potentially leads to different results because the underlying data has changed.
In the cloned project, the constraints will still appear in the interface, and by default they are rescaled for display purposes only. They are not used in the optimization since the preserved Lagrange multipliers take precedence over them.
The following table summarizes how each cloning method affects the new optimization.
Method | Same | Rescale | Score |
Extensive constraints | No change | Scaled by number of rows | Scaled by number of rows, but not used |
Are extensive constraints satisfied? | Guaranteed
| Guaranteed | Not guaranteed |
Interdependence | Yes, the optimized price depends on the dataset distribution | Yes, the optimized price depends on the dataset distribution | No, each optimization is done independently |
Computing the Optim block in the cloned project | An optimization problem is solved. The user is warned if unfeasible. | An optimization problem is solved. The user is warned if unfeasible. | Applies the λ found in the original project. Always succeeds. |