Neuro-Symbolic Behavior Trees (NSBTs) and Their Verification

Serena S. Serbinowska, Diego Manzanas Lopez, Dung Thuy Nguyen, Taylor T. Johnson
Proceedings of the International Conference on Neuro-symbolic Systems, PMLR 288:409-423, 2025.

Abstract

Neural networks have proven to be incredibly powerful and useful in a variety of domains, but are also often opaque and difficult to reason about. This is undesirable in safety-critical systems. An approach to help mitigate this is to utilize a neuro-symbolic approach that combines the power of neural networks and symbolic structures. In this paper, we present Neuro-Symbolic Behavior Trees (NSBTs). NSBTs are behavior trees that utilize neural networks. We provide several examples of NSBTs, including grid-world examples and a representation of a portion of ACAS Xu, an aircraft collision avoidance system. The grid world example considers over 6 million input states for the neural network, while the ACAS Xu example features 5 networks, each with 6 layers of 50 neurons. Additionally, we implemented support for NSBTs in our BehaVerify software tool, and verify certain safety and liveness properties for these NSBTs. Our verification approach also demonstrates how future improvements could be made using existing neural network verification techniques.

Cite this Paper

BibTeX
@InProceedings{pmlr-v288-serbinowska25a, title = {Neuro-Symbolic Behavior Trees (NSBTs) and Their Verification}, author = {Serbinowska, Serena S. and Manzanas Lopez, Diego and Nguyen, Dung Thuy and Johnson, Taylor T.}, booktitle = {Proceedings of the International Conference on Neuro-symbolic Systems}, pages = {409--423}, year = {2025}, editor = {Pappas, George and Ravikumar, Pradeep and Seshia, Sanjit A.}, volume = {288}, series = {Proceedings of Machine Learning Research}, month = {28--30 May}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v288/main/assets/serbinowska25a/serbinowska25a.pdf}, url = {https://proceedings.mlr.press/v288/serbinowska25a.html}, abstract = {Neural networks have proven to be incredibly powerful and useful in a variety of domains, but are also often opaque and difficult to reason about. This is undesirable in safety-critical systems. An approach to help mitigate this is to utilize a neuro-symbolic approach that combines the power of neural networks and symbolic structures. In this paper, we present Neuro-Symbolic Behavior Trees (NSBTs). NSBTs are behavior trees that utilize neural networks. We provide several examples of NSBTs, including grid-world examples and a representation of a portion of ACAS Xu, an aircraft collision avoidance system. The grid world example considers over 6 million input states for the neural network, while the ACAS Xu example features 5 networks, each with 6 layers of 50 neurons. Additionally, we implemented support for NSBTs in our BehaVerify software tool, and verify certain safety and liveness properties for these NSBTs. Our verification approach also demonstrates how future improvements could be made using existing neural network verification techniques.} }
Endnote
%0 Conference Paper %T Neuro-Symbolic Behavior Trees (NSBTs) and Their Verification %A Serena S. Serbinowska %A Diego Manzanas Lopez %A Dung Thuy Nguyen %A Taylor T. Johnson %B Proceedings of the International Conference on Neuro-symbolic Systems %C Proceedings of Machine Learning Research %D 2025 %E George Pappas %E Pradeep Ravikumar %E Sanjit A. Seshia %F pmlr-v288-serbinowska25a %I PMLR %P 409--423 %U https://proceedings.mlr.press/v288/serbinowska25a.html %V 288 %X Neural networks have proven to be incredibly powerful and useful in a variety of domains, but are also often opaque and difficult to reason about. This is undesirable in safety-critical systems. An approach to help mitigate this is to utilize a neuro-symbolic approach that combines the power of neural networks and symbolic structures. In this paper, we present Neuro-Symbolic Behavior Trees (NSBTs). NSBTs are behavior trees that utilize neural networks. We provide several examples of NSBTs, including grid-world examples and a representation of a portion of ACAS Xu, an aircraft collision avoidance system. The grid world example considers over 6 million input states for the neural network, while the ACAS Xu example features 5 networks, each with 6 layers of 50 neurons. Additionally, we implemented support for NSBTs in our BehaVerify software tool, and verify certain safety and liveness properties for these NSBTs. Our verification approach also demonstrates how future improvements could be made using existing neural network verification techniques.
APA
Serbinowska, S.S., Manzanas Lopez, D., Nguyen, D.T. & Johnson, T.T.. (2025). Neuro-Symbolic Behavior Trees (NSBTs) and Their Verification. Proceedings of the International Conference on Neuro-symbolic Systems, in Proceedings of Machine Learning Research 288:409-423 Available from https://proceedings.mlr.press/v288/serbinowska25a.html.

Related Material