Spatial Deep Learning for Site-Specific Movement Optimization of Aerial Base Stations

Unmanned aerial vehicles (UAVs) can be utilized as aerial base stations (ABSs) to provide wireless connectivity for ground users (GUs) in various emergency scenarios. However, it is a NP-hard problem with exponential complexity in <inline-formula> <tex-math notation="LaTeX">$M$ </tex-math></inline-formula> and <inline-formula> <tex-math notation="LaTeX">$N$ </tex-math></inline-formula>, in order to maximize the coverage rate of <inline-formula> <tex-math notation="LaTeX">$M$ </tex-math></inline-formula> GUs by jointly placing <inline-formula> <tex-math notation="LaTeX">$N$ </tex-math></inline-formula> ABSs with limited coverage range. The problem is further complicated when the coverage range becomes irregular due to site-specific blockages (e.g., buildings) on the air-ground channel, and/or when the GUs are moving. To address the above challenges, we study a multi-ABS movement optimization problem to maximize the average coverage rate of mobile GUs in a site-specific environment. The Spatial Deep Learning with Multi-dimensional Archive of Phenotypic Elites (SDL-ME) algorithm is proposed to tackle this challenging problem by 1) partitioning the complicated ABS movement problem into ABS placement sub-problems each spanning finite time horizon; 2) using an encoder-decoder deep neural network (DNN) as the emulator to capture the spatial correlation of ABSs/GUs and thereby reducing the cost of interaction with the actual environment; 3) employing the emulator to speed up a quality-diversity search for the optimal placement solution; and 4) proposing a planning-exploration-serving scheme for multi-ABS movement coordination. In particular, the locations of ABSs/GUs are converted into grid pattern representations, whose dimension and associated DNN complexity are invariant with arbitrarily large <inline-formula> <tex-math notation="LaTeX">$M$ </tex-math></inline-formula> and/or <inline-formula> <tex-math notation="LaTeX">$N$ </tex-math></inline-formula>. Moreover, the virtual emulator-planning combined with the actual site-deployment effectively compensates for the prediction errors due to model approximation. Numerical results demonstrate that the proposed approach significantly outperforms the benchmark Deep Reinforcement Learning (DRL)-based method and other two baselines in terms of average coverage rate, training time and/or sample efficiency. Moreover, with one-time training, our proposed method can be applied in scenarios where the number of ABSs/GUs dynamically changes on site and/or with different/varying GU speeds, which is thus more robust and flexible compared with conventional DRL-based methods.