Loading…

One map per task · Dot: one submission · Line: run order · Hollow dot: first submission · Ringed dot: final pick · Distances compare within a map only

Method

Each dot is placed by its model configuration: the first two principal components of the three agents' designs on that task, so distances compare within a map and not across maps. Shared clusters are found once per task on the three agents' designs together (affinity propagation on the embeddings, as in the blog post), so a cluster number means the same group for every agent on that task. Clusters counts the shared clusters an agent reaches within its first K designs, where K is the smallest design count of the three agents on that task.