Atopological Optimization Through Reinforcement Learning: An Electromagnetic Case Study

Over the past decade, Reinforcement Learning has been investigated for tackling optimization challenges. Specifi-cally, Reinforcement Learning algorithms have been exploited as optimization algorithms in intricate black-box inverse problems, determining the optimal values of the manipulated variables to achieve a specific target. The Reinforcement Learning algorithms train a meta-model, called agent, by exploring the problem space. Then, the agent is used to determine the best values of the handled variables that match the objective. In the literature, the agent is usually trained over a single objective target. Just like other optimization algorithms, if the objective target changes, it is necessary to retrain the agent. In this article, Deep Reinforcement Learning is proposed as an optimization tool capable to learn and optimize the control variables to meet different targets with a single training process. The approach involves the training of a set of Artificial Neural Networks to associate different optimization targets with the corresponding control variable values. To assess its efficacy, the methodology is applied to a case study involving the topology optimization of a 2-D axisymmetric coil using an electromagnetic finite element method. The results highlight comparable performances with an advanced data-driven global optimization algorithm and superior performances than gradient descent, genetic, and particle swarm algorithms for different objectives.