Mahsa Asadi

Machine learning researcher

My research focuses on reinforcement learning, bandits, and federated learning.
Most recently, I worked for three years as a postdoctoral researcher at Aalto University, supervised by Samuel Kaski (human-in-the-loop decision making) and Alex Jung (bandits and GTVMin in machine learning).
I completed my PhD under the supervision of Aurelien Bellet and Marc Tommasi in collaboration with Odalric Maillard, researching structure in online and federated learning problems. Go to my Professional Background

Profession

Postdoctoral Researcher2025 – Present

Machine Learning Research Group, Aalto University, Finland

Supervisor: Alex Jung

Postdoctoral Researcher2023 – 2025

Probabilistic Machine Learning (PML) Group, Aalto University, Finland

Supervisor: Samuel Kaski

Education

Ph.D. in Computer ScienceOct 2018 – Nov 2022

MAGNET team, Inria Lille – Nord Europe & University of Lille, France

Supervisors: Aurélien Bellet, Marc Tommasi, Odalric Maillard

M.Sc. in Artificial Intelligence2013 – 2016

Shiraz University, Iran

GPA: 18.7/20 · Ranked 3^rd

Supervisor: Sattar Hashemi · Advisor: Haitham Bou-Ammar (Princeton University)

Other Education & Research Positions (click to expand)

B.Sc. in Computer Engineering2007 – 2011

Shiraz University, Iran

GPA: 16.50/20 · Ranked 4^th

Diploma in Mathematics and Physics2003 – 2007

NODET — National School for Exceptional Talents, Shiraz, Iran

Summer Schools

Machine Learning Summer School (MLSS)

July 2019

Transylvanian Machine Learning Summer School (TMLSS)

July 2018

Earlier Research Positions

2017

Research Internship · Inria Lille – Nord Europe, SequeL team (4 months)

2016

Shiraz University Center of Intelligent Vision and Image Processing (6 months)

2014 – 2016

Shiraz University Machine Learning Lab

2013 – 2014

Shiraz University Image Processing and Vision Lab

Research

My research bridges reinforcement learning and human modeling; specifically, how learning agents can improve on their decision making via interacting with a human and receiving feedback in various forms. I am interested in both the theoretical foundations of RL and human modeling leading to novel algorithms with the goal of building systems where human-AI interaction is principled and efficient in terms of improving on sample complexity and regret bounds.

Projects

Human-in-the-Loop Decision Making

Structured Bandits with Human Assistance Active Mar 2025 – Present
Leveraging structure in stochastic multi-armed bandit problems to optimize the use of human feedback. The human is modeled as a bounded-rational oracle capable of detecting arms deviating beyond a threshold from the optimal. The goal is to balance exploration and querying to efficiently eliminate suboptimal arms and minimize regret. Current focus: bandits with Lipschitz structure.

BanditsHuman FeedbackLipschitzRegret Minimization

Human-Guided Near-Optimal Regret Bounds Completed Nov 2024 – Mar 2025
Investigated whether side information from a bounded human oracle can improve theoretical guarantees of reinforcement learning. Focused on Upper Confidence Reinforcement Learning (UCRL): proposed an initial realistic user model and improved the regret guarantee of UCRL, providing a theoretical bound on the gain achievable through human assistance.

UCRLHuman OracleRegret BoundsTheory

Lowering Cognitive Burden in RLHF for Drug Design Implemented Summer 2023 – Summer 2024
Proposed preferential feedback types to reduce cognitive burden on human annotators, leveraging the insight that humans find it easier to compare two instances than to evaluate a single one. Introduced a user model to effectively learn the reward using these new query types.

RLHFDrug DesignCognitive LoadReward Learning