Reinforcement Learning for Radar Waveform Optimization

Recently, it has been shown that reinforcement learning (RL) is able to solve decision-based problems through a series of action-observation-reward cycles. In this paper, we pose the problem of constrained waveform optimization as a sequential decision problem and show how it can be solved by an RL agent. The proposed RL-based method is an alternative to mix-integer optimization, evolutionary algorithms, and Bayesian optimization, which is capable of dealing directly with a variable parameter space dimension while considering designs with different processing algorithms in the (optimization) loop. To illustrate the effectiveness of the proposed method, we demonstrate the optimization of an agent's policy capable of defining the number of pulses as well as their duration and modulation parameters of radar waveform while optimizing an user-defined figure of merit.