Learning-Based Maximum Likelihood Estimator for Angle-of-Arrival Localization

The estimation of target positions from angle-of-arrival (AOA) measurements has been extensively researched, and various estimators have been proposed to tackle this challenge. Among these, the maximum likelihood estimator (MLE) is notable for its well-recognized properties, including asymptotic unbiasedness and efficiency. However, traditional MLEs, such as the Gauss-Newton algorithm, often encounter difficulties due to the need for a first-order linearization step in computing the Jacobian matrix. This requirement introduces the potential for significant errors and convergence issues, especially in highly nonlinear systems. To overcome this limitation, this paper introduces a learning framework to address the maximum likelihood estimation problem, where the iterative increments are treated as the output of the agent's actions. Building upon this framework, we develop a learning-based MLE. Comprehensive numerical simulation results demonstrate the effectiveness and superiority of our approach. First, it effectively resolves convergence issues associated with linearization in traditional MLEs. Second, it exhibits robust adaptability by successfully solving both two-dimensional and three-dimensional AOA localization problems. Last, the proposed method significantly enhances localization accuracy compared to existing estimators.