Configuration CubeΒΆ
The configuration cube displays configurations and their score on the objective. Naturally, the space of configurations may be very high dimensional and so this plugins allows you to investigate the configurations through 2D or 3D slices of the configuration space, plotting the values of hyperparameters as points in this slice of the space.
Objective: Here you can select the objective you wish to analyze for. The score for an objective will be communicated through the color of a point with an associated heatmap.
Budget: Here you can specify a filter for which budgets you wish to view. The Combined budget will show all configurations trialed but only show their scores on the highest budget if it was evaluated under multiple different budgets.
Number of Configurations: This slider will allow you to select the number of configurations to display, where the configurations are ordered by they when they were trialed. You can move this along to see how the optimizer chose to evaluate certain hyperparamters over time.
Hyperparameters: Here you can select the slice with which to look into the configuration space. Selecting only one hyperparameter gives you a very basic 1D view, 2 Hyperparameters a 2d grid and 3 hyperparamters a 3D cube. If multiple configurations map to the same point then the one with the highest score for the chosen objective is displayed. This is particular evident in the 1D case where you may have many configurations use the same hyperparameter value.
The Configuration Cube is helpful to visually inspect and answer the following questions:
Does a certain hyperparameter have high influence over the final objectives score when compared to another?
Does the optimizer tend to prefer one value of a hyperparameter over time?
How do hyperparamters correlate with respect to the given objective
