Abstract: System and method for training inductive logic programming enhanced deep belief network models for discrete optimization are disclosed. The system initializes (i) a dataset comprising values and (ii) a pre-defined threshold, partitions the values into a first set and a second set based on the pre-defined threshold. Using Inductive Logic Programming (ILP) engine and a domain knowledge associated with the dataset, a machine learning model is constructed on the first set and the second set to obtain Boolean features, and using the Boolean features that are being appended to the dataset, a deep belief network (DBN) model is trained to identify an optimal set of values between the first set and the second set. Using the trained DBN model, the optimal set of values are sampled to generate samples. The pre-defined threshold is adjusted based on the generated samples, and the steps are repeated to obtain optimal samples.