Layout design using neural networks and Markov random fields

We represent the relations between spatially distributed objects in a design problem by conditional distributions. They define the probability of the different objects in a location as a function of the configuration in the local neighborhood. The conditional distributions induce a Markov random field (MRF) and are estimated by feedforward neural networks from observed configurations. A design for a new floorplan and new marginal constraints is generated by stochastic simulation of the MRF.