I know I'm commenting on an old post here, but I got stuck on this puzzle, and in the current version, this solution does not work BUT it did lead me to the right path.

At the opening I had an Any/Any split. This allowed me to parallel process the slower Perceptron sorter.

Triangles went straight into the output stream.
Other shapes went into a decision tree, with Red going to the output and Blue going to the trash.

explanation:


When I tried this the first time, my solution used another decision tree. I was first sorting on red/green, then red/blue and green/blue. this increased my accuracy to 80%+, but was just a little too slow for gold. I realized only 60% accuracy was necessary, so I could be a little less selective about the colors.

I noted the input contains more blue than green. It seems a decision tree detecting the more frequent color results in greater throughput; this is why I chose a red/blue decision tree. For the curious, beginning with a red/green decision tree results in a throughput of 9.8s. My solution above runs in 8.9s and has a little over 60% accuracy.