Cyber-physical control systems must satisfy strict constrains on the magnitude of control actions and the rate of signal change, while simultaneously balancing multiple, often conflicting, control objectives. In multi-objective optimization (MOO) problems, the optimizer naturally produces a set of competing control candidates. However, in closed-loop operation, a single implementable control action must be selected at each sampling instant. In practice, this selection is typically performed through fixed scalarization rules, which can lead to performance degradation under changing operating conditions in the case of non-convex Pareto fronts or poorly scalable objectives. This work aims to improve the quality of control based on MOO through a decision support system that accounts for the structure: the proposed architecture decouples online Pareto-front construction from action selection. A heuristic multi-objective optimization algorithm based on the NSGA-II genetic algorithm generates feasible control sequences under magnitude and rate constraints, while a learned selector, denoted as Learned Pareto Filter (LPF), evaluates these candidates using structural features that account for both individual characteristics of the solutions and their interrelations within the Pareto set. The method is evaluated on two benchmark problems. First, a system with an analytically known Pareto front structure is used. Second, the approach is validated on a nonlinear boiler-turbine load-changing scenario, widely used in constrained nonlinear control studies. The results demonstrate that LPF-based selection produces smoother control actions, improved tracking performance, and greater robustness to regime changes compared to objective-based scalarization methods. The findings confirm that, in Pareto-based control, closed-loop quality depends critically on the selection mechanism rather than solely on the optimizer generating the Pareto set.