Theoretical and Experimental Results for Planning with Learned Binarized Neural Network Transition Models

Abstract

We study planning problems where the transition function is described by a learned binarized neural network (BNN). Theoretically, we show that feasible planning with a learned BNN model is NP-complete, and present two new constraint programming models of this task as a mathematical optimization problem. Experimentally, we run solvers for constraint programming, weighted partial maximum satisfiability, 0–1 integer programming, and pseudo-Boolean optimization, and observe that the pseudo-Boolean solver outperforms previous approaches by one to two orders of magnitude. We also investigate symmetry handling for planning problems with learned BNNs over long horizons. While the results here are less clear-cut, we see that exploiting symmetries can sometimes reduce the running time of the pseudo-Boolean solver by up to three orders of magnitude.