Abundant records now link organizational contexts, actions, and outcomes. Evolutionary Surrogate-Assisted Prescription (ESP) converts such data into trustworthy policies through a two-stage neuro-symbolic pipeline: a neural network Predictor surrogate is trained first using supervised learning, after which an interpretable Prescriptor is evolved against it using the EVOTER rule grammar. Decoupling prediction from prescription yields high sample-efficiency, low on-line risk, and explicit regularization, while the resulting rule sets remain compact and auditable. Across diverse, safety-critical domains, ESP attains accuracy on par with—or exceeding—neural network models, yet retains full transparency, establishing a robust platform for large-scale, trustworthy decision optimization.

PDF

In Proceedings of Machine Learning Research vol 284:1–3, 2025: 19th Conference on Neurosymbolic Learning and Reasoning, 2025.

@inproceedings{shahrzad:nesy25, title={Evolutionary Surrogate-Assisted Prescription: Neuro-Symbolic Framework for Trustworthy Decisioning}, author={Hormoz Shahrzad and Risto Miikkulainen}, booktitle={Proceedings of Machine Learning Research vol 284:1–3, 2025: 19th Conference on Neurosymbolic Learning and Reasoning}, month={ }, url="http://nn.cs.utexas.edu/?shahrzad:nesy25", year={2025} }

Poster