A Decision-Tree-Based Multi-objective Estimation of Distribution Algorithm

A new decision-tree-based multi-objective estimation of distribution algorithm (DT-MEDA) for optimization problems with continuous variables is developed. Decision-tree-based probabilistic models are used to encode conditional dependencies among variables in DT-MEDA. By building and sampling the probabilistic models, the algorithm reproduces the genetic information of the next generation. Incorporating this reproduction mechanism together with the ranking method and the truncated selection, DT-MEDA can approximate the Pareto front. Polynomial mutation operator is used to enhance exploration and maintain diversities in the populations. Furthermore, DT-MEDA adopts a procedure to eliminate a solution with smallest crowding distance at a time in the truncated selection, so that it can obtain a well spread solution set. The performance of the proposed algorithm is evaluated on four biobjective test problems and metrics from literature. Simulation results show that the proposed approach is competitive with NSGA-II and DT-MEDA is a general and effective method for multi-objective optimization.