A new method for linkage learning in the ECGA

The ECGA is a competent Genetic Algorithm that uses a probabilistic model to learn the linkage among variables and then uses this information to solve hard problems using polynomial resources. However, in order to detect the linkage, the ECGA needs to perform a quadratic number of evaluations of a metric called CCC, a time consuming process. This paper presents ClusterMI, a new method for linkage detection in the ECGA. ClusterMI requires only a linear number of evaluations of the CCC reducing the overall running time of the algorithm. Experiments show that ClusterMI retains ECGA's scalability concerning population size while reducing the running time by $O(m^{0.7})$, resulting in speedups of potentially thousands of times (estimated speedup for a problem with $2^{20}$ bits is 1515).