This paper explores the use of extended grammars in Grammatical Evolution (GE) to solve complex problems while maintaining interpretability. By incorporating external knowledge through custom-defined grammars, the proposed approach aims to generate solutions that are both accurate and explainable. Two new grammars are introduced: one liberal and one strict, which enhance GE by incorporating neural network-inspired structures. These grammars include activation functions commonly used in neural networks and allow the construction of models that mimic neural network behavior without the need for training. Experiments conducted on various datasets demonstrate that the use of these grammars leads to improved robustness, accuracy, and model explainability compared to traditional methods. The results suggest that GE can be a powerful tool for building interpretable models that challenge the dominance of black-box approaches, such as neural networks, making it a promising candidate for Explainable AI (XAI) applications. Furthermore, the reproducibility of the experiments is ensured through publicly available Jupyter notebooks and datasets, allowing for the verification of results and further exploration.