Embedding Normative Reasoning into Neural Symbolic Systems

Normative systems are dynamic systems because their rules can change over time. Considering this problem, we propose a neuralsymbolic approach to provide agents the instruments to reason about and learn norms in a dynamic environment. We propose a variant of d’Avila Garcez et al. Connectionist Inductive Learning and Logic Programming(CILP) System to embed Input/Output logic normative rules into a feed-forward neural network. The resulting system called Normative-CILP(NCILP) shows how neural networks can cope with some of the underpinnings of normative reasoning: permissions, dilemmas, exceptions and contrary to duty problems. We have applied our approach in a simplified RoboCup environment, using the N-CILP simulator that we have developed. In the concluding part of the paper, we provide some of the results obtained in the experiments.