With the Antminer+ algoritm David Martens, Manu De Backer, Raf Haesen,
Bart Baesens and Tom Holvoet at the University of Leuven, Belgium try to classify observations by simulating the behaviour of foraging ants. The paths these fake ants follow lead from nowhere to a decision rule. Each step connects two rules, a rule being something like ‘armlength less then 67 cm’ and the total path is the resulting combined if-statement. If the resulting combined rule is an improvement in the solution space feromones are added to the path so that the probability that the same steps are reused rises. Unused steps see their feromones evaporate. Exactly the way ants searchfor food.
Since each possible rule is connected to each other possible rule, and each rule is a possible value of a categorical variable, the connection space increases exponentially with the number of variables. So the principle is nice, but the usability, e.g. for real world datamining purposes, is extremely limited. A dataminer in a financial institution rather uses 1000-odd variables. If you transform 1000 variables in 5 categories per variable, this gives you (5000 times 5000 minus 5000)/2 connections = 12547500 connections or pheromone levels.
More scalable is the solution by Michelle Galea and Qiang Shen in a chapter in Ajith Abraham, Crina Grosan, Vitorino Ramos : Swarm Intelligence in Data Mining . Although they call it “Simultaneous Ant Colony Optimization Algorithms for Learning Linguistic Fuzzy Rules” what they really describe are locusts. They hop from one rule to another, so on their way there are no pheromones but thin air. In stead the places where they land can contain different pheromone levels, influencing their choice either to jump further or to stay (meaning to grasp the rule). Here the number of pheromone levels is exactly the same as the number of rules (5000 in my previous business example).
The drawback of this locust method is that there is no direct interaction between the different rules, because there is no path connecting them. A locust hops trough the sky and can land anywhere. But the usage of a swarm of locusts should easily cope with this disadvantage.
I like locusts. For together with this scalability the authors show the possibility of learning multiple rules simultaneously. And there is also the fact that they use fuzzy rules.
Very promising !