Multi-objective optimal control of nonlinear processes using reinforcement learning with adaptive weighting

This work proposes a reinforcement learning (RL) framework for solving multi-objective optimal control problems for nonlinear systems. The non-dominated sorting genetic algorithm II (NSGA-II) is integrated into the RL framework to first compute a set of Pareto-optimal solutions that will be used to design adaptive weights. Unlike conventional fixed-weight methods, the proposed framework dynamically adjusts the weights of multiple objectives in response to varying process conditions to improve the trade-off among multiple objectives. Policy iteration is used to optimize control policies under the adaptive weights. The proposed framework is applied to a nonlinear chemical process to demonstrate its effectiveness and superiority over fixed-weight learning methods.