Neural Programs: Linking Probabilistic and Differential Programming

Abstract

We review neural programs (NP), a differential programming paradigm for adaptive controllers we proposed in [1], with close ties to probabilistic programming and machine learning. NPs use a smooth version of if and while statements, where the original boolean condition is interpreted as the mean of a sigmoidal distribution. This way, one can use continuous optimization techniques, to learn the distribution’s variance and the condition’s parameters, such that the NP together with the plant it controls reproduce the behavior specified by a given data set. Given the strong connection between their branching conditions and artificial neurons, NPs allow to conveniently write robust and adaptive controllers, in form of recurrent neural networks (RNNs). We illustrate the utility of NPs on two case studies: parallel parking and tap withdrawal.