Agent Based Evolutionary Algorithms Applied To Constrained Multi-Objective Optimization And Decision-Making

The objectives of multi-objective optimization problems (MOP) generally conflict with each other, which leads to a set of feasible optimal solutions. Additionally, the presence of the constraints might cause non-continuity over the solution space, resulting in difficulty in solution searching. With parallel searching features available, evolutionary algorithms (EA) are suitable to deal with MOP problems. However, challenges stemmed from constraints still remain in convergence speed and solution quality of EA. Thanks to characteristics of autonomy, coordination and self-organizing, agents are recognized to help improve the performances of evolutionary algorithms. Inspired by these observations, in-depth investigations relevant are carried out in this thesis, including following issues.1. Philosophy and metrics of agent are involved into EA, where fundamental properties of agents are specified to make EA more intelligent. The agents are distributed over the feasible space, both performing self-study and receiving relevant information from environment to accelerate searching for Pareto solutions.2. Some operations are proposed regarding the constraints. Violation degrees of the constraints are considered as additive objectives able to influence the energy of agents. Two external archives, optimal solution set and feasible optimal solution set, are available to maintain the diversity of the population and ensure the optimal solutions are preserved during optimization. Climbing operators are suggested to promote both candidate solutions and agents with small violation degrees, achieving feasible optimal solutions efficiently.3. EA based optimization algorithms are combined with multi-objective decision making mechanism, in which partial attributes associated with agents and corresponding optimization operations are modified. Related studies and simulations on the availability of involvement of prior priority and priority evolution approaches are performed, showing satisfied results.4. The proposed agent based evolutionary algorithms are applied to welded beam design and speed reducer design problems separately. Experimental results show that the algorithms not only guarantee solution qualities and distributions but also converge quickly to the optimal fronts, demonstrating the efficiencies.