Formula Neural Networks as Action Policies in Safety-Critical Planning: An Empirical Evaluation

Abstract

Formula neural networks (FNNs) are a recently proposed alternative to neural networks (NNs), with Boolean neuron values and hidden-layer nodes defined as propositional formulas over the previous layer. FNNs have been successfully applied to classification and, compared to standard NNs, offer greater interpretability. Here we show that they are also a promising alternative for learning action policies and that they can be more amenable to formal verification. We consider planning under uncertainty, with the objective to reach the goal while avoiding unsafe states. We adapt a recently proposed NN policy safety verification algorithm to FNNs. Our experiments show that, compared to NNs, FNNs can often learn competitive policies while being much faster to verify.