From: A review of estimation of distribution algorithms in bioinformatics
Statistical order | Advantages | Disadvantages | Examples |
---|---|---|---|
Univariate | Simplest and fastest | Ignore feature dependencies | PBIL (Baluja, 1994) |
Suited for high cardinality problems | Bad performance for deceptive problems | UMDA (Mühlenbein and Paaß, 1996) | |
Scalable | cGA (Harik et al., 1999) | ||
Bivariate (statistics of order two) | Able to represent low order dependencies | Possibly ignore some feature dependencies | MIMIC (De Bonet et al., 1996) |
Suited for many problems | Slower than univariate EDAs | Dependency trees EDA (Baluja and Davies, 1997) BMDA (Pelikan and Mühlenbein, 1999) | |
Graphically inquire the induced models | Tree-EDA/Mixture of distributions EDA (Santana et al., 1999) | ||
Multivariate (statistics of order greater than two) | Parameter learning (only interaction model parameters) | ||
Suited for problems with known underlying model | Possibly ignore complex feature dependencies | FDA (Mühlenbein et al., 1999) | |
Higher memory requirements than bivariate | Markov network-based EDA (Shakya and McCall, 2007) | ||
Structure+parameter learning (interaction model & parameters of the model) | |||
Maximum power of generalization | Highest computation time | EcGA (Harik et al., 1999) | |
Flexibility to introduce user dependencies | Highest memory requirements | EBNA (Etxeberria and Larrañaga, 1999) | |
Online study of the induced dependencies | BOA/hBOA (Pelikan et al., 1999, 2005) | ||
Dependency networks EDA (Gámez et al., 2007) |