A Dynamic Metaheuristic Network for Numerical Multi-objective Optimization

This research work proposes a dynamic metaheuristic network that is a layered interconnection of a number of multi-objective optimization (MOO) algorithms. Each node of the network corresponds to a MOO metaheuristic and interconnections between the nodes represent the flow of subpopulation elements in a feed-forward direction. The proposed method runs in consecutive sessions such that a session starts with the assignment of subpopulations to each of the individual nodes, proceeds with execution of node metaheuristics within their algorithmic framework and ends with feeding the improved subpopulations to the connected forward nodes. The network architecture is dynamic in the sense that nodes change their layers at the end of each session. At the end of each session, elements of the improved subpopulations are fed forward to nodes in subsequent layers which update their own subpopulations using uniform random sampling. The proposed method is evaluated on CEC2009, ZDT, DTLZ, WFG benchmarks and several real-world MOO problems using the experimental framework described for these problem instances. Comparative evaluations against a large set of state-of-the-art algorithms exhibited that the proposed method with its novel dynamic network architecture and subpopulation assignment strategy is promising both in the quality of the extracted Pareto fronts and in leading future research on ensembles of MOO algorithms.