Towards Neural-Symbolic Learning to support Human-Agent Operations

This paper investigates neural-symbolic policy learning for information fusion in distributed human-agent operations. The architecture integrates a pre-trained neural network for feature extraction, with a state-of-the-art symbolic Inductive Logic Programming (ILP) system to learn policies, expressed as a set of logical rules. We firstly outline the challenge of policy learning within a military environment, by investigating the accuracy and confidence of neural network predictions given data outside the training distribution. Secondly, we introduce a neural-symbolic integration for policy learning and demonstrate that the symbolic ILP component, when considering the length of the learned policy rules, can generalise and learn a robust policy despite unstructured data observed at policy learning time originating from a different distribution than observed during training.