Abstract: Fully covering a metallic target with a radar absorbing material is the most effective way to reduce its radar signature although not always practically viable especially for large surfaces. Conversely, randomly positioning radarabsorbing materials on a portion of the target may inadvertently worsen its Radar Cross Section (RCS) precisely in the region of space where we aim to reduce it. To address such constraints, a Reinforcement Learning (RL) approach for the optimal positioning of absorbing tiles on a sub-region of a planar metallic object to reduce its RCS is presented in this paper. By dividing the plate into 9 square sub-regions, we developed a model to predict the RCS response as a function of the incidence angles of the incoming wave and the distribution of the absorbing tiles. Specifically, the proposed model leverages on a Reinforcement Learning algorithm belonging to the Deep Q-Learning Network (DQN) family. The network's agent takes as inputs the plane wave incidence angles (θ and φ) and the number of tiles, producing a tile distribution that minimizes the bistatic RCS in a fixed scattering region. The agent's performance was evaluated by exploiting a collection of incidence directions, achieving a positive reward percentage exceeding 70%.
