• genetic algorithm
• minihack
• nethack
• nle
• reinforcement learning
• rule-based system

This work studies two novel RL benchmarks based on the NetHack videogame: the NetHack Learning Environment and MiniHack. Given its complexity, NetHack represents an open challenge in the field, given that current neural models are able to achieve like a human beginner.
First, we will use pure neural approaches, reaching state-of-the-art results, and then we will integrate them with a rule-based system, to inject prior knowledge and improve the performance.
We will show that the usage of hybrid architectures has beneficial effects both on the training phase and on the performance of the agent, proving that neuro-symbolic solutions must be further investigated to obtain better, more sustainable models.