Markov models and neural networks are widely used in systems tasked with generating natural and meaningful sequences. Generating high-quality sequences often requires the system to impose structure on the sequences via user-defined control constraints. These models are typically not compatible with control constraints. Work has been done to combinenon-hidden Markov models with constraints; however, this approach has the problem of diminishing solution space sizes for increasing Markov orders or constraint complexity. For neural networks, the anticipation-Recurrent Neural Network (anticipation-RNN) allows control constraints but is limited in what kind of constraints work effectively. We propose an efficient method to apply control constraints to a hidden Markov model that, like the non-hidden variant, 1) guarantees sequences generated satisfy constraints and2) the statistical distribution of the constrained model is the same as the original model.The proposed model satisfies the control constraint requirement and avoids the problem of diminishing solution space sizes afforded by the abstraction introduced by the hidden states. Keywords: Sequence Generation, Markov Models, Constraint Satisfaction, ConstrainedHidden Markov Process, Anticipation-RNN |