Matteo Pirotta
Technische Universität Darmstadt, Politecnico di Milano, Meta (Israel)
Papers
7
Total Citations
143
H-Index
5
About
Matteo Pirotta is a reinforcement learning researcher whose work sits at the intersection of policy optimization, safe learning, and multi-objective decision-making — areas with profound implications for real-world applications in robotics, healthcare, and beyond. His most influential contribution, a 2016 paper on multi-objective reinforcement learning through Pareto manifold approximation (51 citations), addresses the fundamental challenge of balancing conflicting objectives, a problem central to practical control systems. Complementing this, his early work on adaptive step-size methods for policy gradient algorithms (48 citations) helped establish more principled and efficient training procedures for continuous control tasks. A recurring theme across Pirotta's research is safety: his work on smoothing policies and safe policy gradients tackles the critical challenge of ensuring that learning agents do not behave dangerously during the training process itself — a major barrier to deploying reinforcement learning in sensitive domains. His investigations into conservative exploration in bandits further reflect this commitment to responsible learning under uncertainty. Notably, Pirotta has also explored the clinical frontier, applying reinforcement learning to functional electrical stimulation for stroke rehabilitation, demonstrating a genuine drive to translate theoretical advances into tangible human benefit.
Research Focus
Key Achievements
Top Papers
- 1
- 2Adaptive Step-Size for Policy Gradient Methods48 citations · 2013
- 3Smoothing policies and safe policy gradients20 citations · 2022
- 4Improved Algorithms for Conservative Exploration in Bandits12 citations · 2020
- 5Smoothing Policies and Safe Policy Gradients6 citations · 2019
- 6
- 7Fitted policy search2 citations · 2011