Prescriptive information fusion

Enterprise big-data analytics requires data from diverse sources to be fused and harmonized after which it becomes useful for mining interesting patterns as well as to make predictions. However, neither are all patterns equally insightful nor are predictions of much value unless they can support decisions: Prescriptive rather than mere predictive analytics is needed, which involves optimization in addition to traditional predictive modeling. Further, because of the paucity of real-data covering a large enough space of decisions, simulations based on a theory of the world can also be used to augment real data while learning statistical models for prescriptive purposes. In this paper we present a unified Bayesian framework for prescriptive information fusion that formally models the iterative fusion of information from simulation, statistical as well as optimization models, over and above the fusion of information from multiple data sources. We motivate our framework with diverse real-life applications including warranty provisioning, the computational design of products or manufacturing-processes, and the optimal pricing or promotion of consumer goods. We also compare our approach with reinforcement learning, as well as other combinations of machine-learning, simulation and optimization.