Automatic design of microwave filters by using relational induction neural networks

Abstract

AbstractAutomatic design of microwave filter based on artificial intelligence is recognized as an alternative to conventional approach, due to the complexity of the structure parameters and the need of enormous solution space. In this paper, an unsupervised, end‐to‐end training framework, which is named Relational Induction Neural Network (RINN) is proposed. RINN discretizes the model, and makes cluster analysis on each action of geometrical parameters. Deep reinforcement learning (DRL) neural network is applied to design the filter according to reward function. Several second‐order filters with different performance and one sixth‐order filter are designed with RINN. The results show that RINN can automatically induce the relationship between microwave device structures without any prior knowledge through training or learning. This research crosses the gap between DRL and microwave filter, and can also be extended to other microwave device, antenna and other related fields.