We address the problem of efficient and unobstructed surveillance orcommunication in complex environments. On one hand, one wishes to use a minimalnumber of sensors to cover the environment. On the other hand, it is oftenimportant to consider solutions that are robust against sensor failure oradversarial attacks. This paper addresses these challenges of designing minimalsensor sets that achieve multi-coverage constraints -- every point in theenvironment is covered by a prescribed number of sensors. We propose a greedyalgorithm to achieve the objective. Further, we explore deep learningtechniques to accelerate the evaluation of the objective function formulated inthe greedy algorithm. The training of the neural network reveals that thegeometric properties of the data significantly impact the network'sperformance, particularly at the end stage. By taking into account theseproperties, we discuss the differences in using greedy and $\epsilon$-greedyalgorithms to generate data and their impact on the robustness of the network.