The relative disposition of individual telescopes in the ground is one of the important factors in optimising the performance of a stereoscopic array of imaging atmospheric Cherenkov telescopes (IACTs). Following previous attempts at an automated survey of the broad parameter space involved using evolutionary algorithms, in this paper we will present a novel approach to optimising the array geometry based on deep learning techniques. The focus of this initial work will be to test the algorithmic approach and will be based on a simplified toy model of the array. Despite being simplified, the model heuristics aims to capture the principal array performance features relevant for the layout optimisation. Our final goal is to create an algorithm capable of scanning the large parameter space involved in the design of a large stereoscopic array of IACTs to assist optimisation of the array geometry (in face of external constraints and multiple performance objectives). The use of simple heuristics precludes direct comparison to existing real-world experiments, but the analysis is internally consistent and gives insight as to the potential of the technique. Deep learning techniques are being increasingly applied to tackle a number of problems in the field of Gamma-ray Astronomy, and this work represents a novel, original application of this modern computational technique to the field.