EGNN: A Deep Reinforcement Learning Architecture for Enforcement Heuristics

An increasing amount of research is being directed towards neuro-symbolic computing, combining learning in neural networks with reasoning and explainability via symbolic representations [4]. One subfield of AI where neuro-symbolic methods are a promising alternative for existing symbolic methods is computational argumentation. Much of the theory of computational argumentation is based on the seminal work by Dung [6], in which he introduces abstract argumentation frameworks (AFs) of arguments and attacks, and several acceptability semantics that define which sets of arguments (extensions) can be reasonably accepted. Core computational problems in abstract argumentation are typically solved with handcrafted symbolic methods [1]. However, recently we demonstrated the potential of a deep learning approach by showing that a graph neural network is able to learn to determine almost perfectly which arguments are (part of) an extension [2]. When considering dynamic argumentation a growing research area where the knowledge about attacks between arguments can be incomplete or evolving other types of computational problems arise where neuro-sybmolic methods are still unexplored. In [3] we propose our enforcement graph neural network (EGNN), a learning-based approach to the dynamic argumentation problem of enforcement: given sets of arguments that we (do not) want to accept, how to modify the argumentation framework in such a way that these arguments are (not) accepted, while minimizing the number of changes [5]. Here we demonstrate our implementation of an EGNN.