Isidoro Tamassia

PhD Researcher @ KU Leuven

NeuroSymbolic Reinforcement Learning

About Me

I am a PhD researcher (Oct. 2025-present) at the DTAI group of KU Leuven University, supervised by Prof. Giuseppe Marra. My research interests lie at the intersection of NeuroSymbolic AI and Reinforcement Learning.

I hold an MSc degree in Computer Science from TU Delft University, where I graduated under the supervision of Prof. Wendelin Böhmer [thesis]. During my master's program, I had the opportunity to conduct research in both the theory and applications of Deep Reinforcement Learning [publications].

Previously, I obtained a BSc degree in Computer and Control Engineering from Università di Roma 'La Sapienza'.

Research Field: NeSy + RL

Knowledge compilation in DeepSeaProbLog, a modern NeSy language.

NeuroSymbolic AI

NeuroSymbolic AI (NeSy) is widely considered the 3rd wave of AI. By combining the paradigms of learning and reasoning, NeSy tries to take advantage of the representational power of neural networks as well as the sound formalisms of logic and probabilistic reasoning. Nonetheless, there exist relatively few established frameworks that allow the development of NeSy systems, and their scalability cannot match that of purely neural approaches yet. Current NeSy research at DTAI focuses on bridging this gap by working at both the conceptual and computational level.

Example of a Deep Reinforcement Learning loop.

Reinforcement Learning

Reinforcement Learning (RL) teaches an agent to interact with an environment through positive or negative feedback. Modern RL frameworks employ neural network architectures for improved scalability (DeepRL), allowing them to tackle impressively complex challenges in fields such as robotics or game playing. However, generalization to novel test environments, as well as the explainability of the learned policies, remain open problems and therefore active fields of research within the machine learning community.

Our goal is to develop novel RL methods capable of leveraging modern NeSy frameworks. This could potentially allow more reliable enforcement of several behavioral constraints during the learning and deployment processes, such as safety constraints that are critical in many real-world applications. Moreover, we aim to push the boundaries of DeepRL agents' generalization capabilities and provide new ways of learning explainable RL policies.

Page updated

Google Sites

Report abuse