Leveraging Reward Machines for Efficient Multi-Objective Reinforcement Learning

Panos Aronis, Mehdi Dastani, Roxana Rădulescu, Giovanni Varricchione; 7, 2026.

Presented at the Reinforcement Learning Conference (RLC), Montréal, Quebec, Canada, August 15–17, 2026.

Abstract

Reinforcement Learning (RL) provides a powerful framework for sequential decision-making, typically assuming a single scalar reward function and a Markovian reward structure. However, many real-world problems involve multiple, potentially conflicting objectives and require temporally extended behaviours that fundamentally violate the Markov property. These complexities have separately motivated the study of Multi-Objective Reinforcement Learning (MORL) and Reward Machines (RMs), both of which demand new representations and learning strategies to ensure effective, sample-efficient, and interpretable solutions. To address these challenges, we propose a novel framework, Multi-Objective Reward Machines (MORMs), to handle non-Markovian reward structure and multiple objectives in a unified setting, enabling more principled and sample-efficient learning in complex sequential decision-making tasks.

[abs][pdf]

BibTeX

@article{aronis2026leveraging,
    title={Leveraging Reward Machines for Efficient Multi-Objective Reinforcement Learning},
    author={Panos Aronis and Mehdi Dastani and Roxana Rădulescu and Giovanni Varricchione},
    journal={Reinforcement Learning Journal},
    volume={7},
    pages={},
    year={2026}
}