Representative designs from an evaluated landscape
ClusterMedoid partitions a discrete evaluated Sim::OPT landscape into clusters and selects one representative medoid for each cluster. A medoid is an actual observed design instance minimizing total within-cluster dissimilarity, not an artificial arithmetic centroid.
This is useful when every reported representative should correspond to a realizable, already evaluated design.
Context and problem variables
Variables can be separated into context variables and problem variables. One-level variables are accepted as fixed coordinates and are retained for validation and output while being omitted from distance calculations. Re-embedded landscapes can also declare fixed global levels explicitly.
Similarity
For an ordered discrete variable with L levels, the variable distance is based on the logarithm of the level separation and is converted to a similarity. Variable similarities are combined within context and problem groups by a hybrid arithmetic-geometric measure, controlled by a mixing parameter.
A selected performance column is also placed on a normalized logarithmic scale. The context, problem and performance components are then combined into a final hybrid similarity, with dissimilarity defined as one minus that similarity.
Clustering
Exact k-medoids is available and uses deterministic initial medoids followed by assignment and recomputation. When the requested number of clusters is auto, silhouette information is used in the documented exact path.
For larger problems, the module can use CLARA: exact PAM on samples followed by validation and assignment of the full dataset. In automatic mode the implementation selects the exact route only when the pairwise distance matrix fits within its configured memory budget.