In-region location verification (IRLV) in wireless networks is the problem of deciding if user equipment (UE) is transmitting from inside or outside a specific physical region (e.g., a safe room). The decision process exploits the features of the channel between the UE and a set of network access points (APs). We propose a solution based on machine learning (ML) implemented by a neural network (NN) trained with the channel features (in particular, noisy attenuation values) collected by the APs for various positions both inside and outside the specific region. The output is a decision on the UE position (inside or outside the region). By seeing IRLV as an hypothesis testing problem, we address the optimal positioning of the APs for minimizing either the area under the curve (AUC) of the receiver operating characteristic (ROC) or the cross entropy (CE) between the NN output and ground truth (available during the training). In order to solve the minimization problem we propose a twostage particle swarm optimization (PSO) algorithm. We show that for a long training and a NN with enough neurons the proposed solution achieves the performance of the Neyman-Pearson (N-P) lemma.