It is common knowledge that using directional antennas is often mandatory for Multi-hop ad-hoc wireless networks to provide satisfying quality of service, especially when dealing with an important number of communication nodes [1]. As opposed to their omnidirectional counterpart, directional antennas allow for much more manageable interference patterns: a receiving antenna is not necessarily interfered by nearby emitting antennas as long as this receiving antenna is not directed towards these undesired emission beams. Two nodes then need to steer one of their antennas in the direction of the other node in order to create a network communication link. These two users will then be able to, in turn, emit and receive to and from each other. The scope of this work resides in finding a centralized algorithm to governate these antenna steering decisions for all users to instantaneously provide a valid set of communication links at any time given the positions of each user. The problem that raises is then a geometrical one that implies finding topologies of network links that present satisfying throughput and overall QoS and guarantee instantaneous connectedness i.e. the computed set of links allows any user to reach any other user in a certain number of hops. Building such optimized link topologies makes further tasks, such as routing and scheduling of the network, much simpler and faster. This problem is highly combinatorial and, while it is solvable with traditional Mixed Integer Programming (MIP), it is quite challenging to carry it out in real time. For this purpose, we propose a Deep Neural Network that is trained to imitate valid, solved instances of the problem. We use the Attention mechanism [2] [3] to let nodes exchange information in order to capture interesting patterns and properties that then enable the neural network to generate valid network link topologies, even dealing with unseen sets of users positions.