About
Girolamo Macaluso
PhD Student in Artificial Intelligence
Media Integration and Communication Center, University of Florence, Italy
馃搷 Florence, Italy 路 鉁夛笍 girolamo.macaluso@unifi.it 路 馃敆 LinkedIn 路 馃捇 GitHub 路 馃帗 Scholar
My research is dedicated to making Reinforcement Learning more efficient and scalable. I began by exploring offline and hybrid offline鈫抩nline RL methods to maximize learning from static datasets, then developed techniques to cut the computational and memory costs of state-of-the-art algorithms. I鈥檓 now advancing continual RL frameworks that adapt to evolving environments and applying RL to fine-tune diffusion models.
Education
-
PhD in Artificial Intelligence
Nov 2023鈥揚resent 路 University of Florence
Topic: Efficient Reinforcement Learning for game development. -
MSc in Artificial Intelligence
Sep 2021鈥揓ul 2023 路 University of Florence
Thesis: Enhancing Offline RL with Generative World Models
Grade: 110 L/110 (GPA 4.0) -
BSc in Computer Engineering
Sep 2018鈥揓ul 2021 路 University of Florence
Experience
Part-time Software Engineer, Maba S.R.L, Florence
Jan 2019鈥揓ul 2023
Designed and delivered end-to-end enhancements to a centralized ERP platform, including database modeling, backend API development, and user-facing interface components to support customer-specific workflows.
Research Interests
- Offline Reinforcement Learning: leverage pre-collected data without online interaction
- Fine-tuning of Diffusion Models: adapt diffusion models to different tasks with reinforcement learning
- Continual Reinforcement Learning: create agents that are capable to adapt to changes of the environment
- Reinforcement Learning for Game Development: accelerate and improve video game development with reinforcement learning
Publications
-
SPEQ: Offline Stabilization Phases for Efficient Q-Learning in High Update-To-Data Ratio RL.
Accepted at Reinforcement Learning Conference 2025. 馃敆Open Review 馃敆GitHub (Code available soon)
Introduces a novel online reinforcement learning framework with periodic offline stabilization that cuts computational requirements and training time while matching or exceeding state-of-the-art performance. -
A Benchmark Environment for Offline RL in Racing Games.
Oral at IEEE Conference on Games 2024. 馃敆Arxiv 馃敆GitHub
Introduces OfflineMania a novel benchmark environment for offline and offline to online RL with challenging data distribution. -
Small Dataset, Big Gains: Enhancing RL by Offline Pre-Training with Model-Based Augmentation.
Oral at AIBSD Workshop, AAAI 2024. 馃敆Arxiv
Proposed a novel technique based on a generative model on for improving offline to online RL performance when dealing with small datasets.
Teaching & Mentoring
Teaching Assistant, University of Florence
Jan 2023鈥揚resent
Interactive C/C++ & Python lessons for 200+ undergraduates.
STEM Outreach Teacher, European STEM Project
Feb 2024鈥揚resent
Introduce students to the field of engineering and explain the challenges of applying AI to real world problems.