Embedding Decision Diagrams into Generative Adversarial Networks

Abstract

Many real-world decision-making problems do not possess a clearly defined objective function, but instead aim to find solutions that capture implicit user preferences. This makes it challenging to directly apply classical optimization technology such as integer programming or constraint programming. Machine learning provides an alternative by learning the agents’ decision-making implicitly via neural networks. However, solutions generated by neural networks often fail to satisfy physical or operational constraints. We propose a hybrid approach, DDGan, that embeds a Decision Diagram (DD) into a Generative Adversarial Network (GAN). In DDGan, the solutions generated from the neural network are filtered through a decision diagram module to ensure feasibility. DDGan thus combines the benefits of machine learning and constraint reasoning. When applied to the problem of schedule generation, we demonstrate that DDGan generates schedules that reflect the agents’ implicit preferences, and better satisfy operational constraints.