When I started my research work in the University of Tokyo, my first topic was the study of Ant Colony Clustering algorithms (ACC). I tried to use some Meta-GA to improve ACC, which was believed at the time to be very sensitive on a number of parameters. It was a fun and exciting research, and I read a slew of very interesting papers. However, eventually I hit a wall on that research, in which the clustering generated by ACC would be too fragmented (one class was broken into many subclasses). I couldn’t solve that problem, so I eventually gave up on this line of research and switched to portfolio optimization instead.

Today, a fellow swarm researcher, Vitorino Ramos sent a link via twitter on a new paper about ACC. The paper was a modeling from ACC to SOM that addressed exactly the “too many clusters” problem I was having. Their key idea was that the ACC could be maping as a sample method for neighborhood clustering. So they kinda got rid of the ants, but solved the problem. It was a bit sad to see someone else solve that problem I couldn’t solve a few years ago, but it was also nice that they cited my ACC paper (even including a figure from it), when describing the problem with traditional ACC.

Got a bit Nostalgic with this…