Neuro-Symbolic AI

Neuro Symbolic AI is very much a hybrid of (good-old-fashioned) symbolic AI methods and “modern” deep learning methods. The main idea of this research direction is to complementary enrich the capabilities of each of the individual research directions with the capabilities of the other. For example, while on the one hand symbolic AI methods, like formal reasoning, satisfiability modulo theory, etc. are very good in explainable reasoning, deep learning methods are not. On the other hand deep learning methods are very good at modeling from data, which is a pain for formal models. So the idea of Neuro-Symbolic methods is to fuse the best of both worlds in hybrid approaches, like using deep learning for modeling but rely on formal reasoning for getting explainable results.

As long as deep learning methods cannot fill the capabilities of symbolic methods, Neuro-Symbolic AI is a viable transitional technology. We are currently working on Neuro-Symbolic approaches, especially in the planning research domain that surpass very conservative approaches, such as action model learning. Our first landmark publication is the “Lazy Approach to Neural Numerical Planning with Control Parameters”. This method integrates learned models into a Planning as Satisfiablity paradigm.

Currently we are working on another publication from a different angle, about integrating formal reasoning and knowledge into Reinforcement Learning. Stay tuned for more…