Publications

• Sorted by Date • Classified by Publication Type • Classified by Research Category •

Objective-Behavior Alignment: Diagnostics for MORL Policy Selection

Antonio Mone, Zuzanna Osika, Florian Felten, Pradeep K. Murukannaiah, Mark Fuge, Frans A Oliehoek, and Luciano Cavalcante Siebert. Objective-Behavior Alignment: Diagnostics for MORL Policy Selection. Transactions on Machine Learning Research, 2026.

Download

pdf [1.6MB]  

Abstract

Real-world decision-making often requires optimizing multiple competing objectives simul- taneously. In reinforcement learning (RL), this is typically addressed by combining reward signals into a single scalar objective via a scalarization function, which can be fragile: small changes in the weights can induce drastically different policies. Multi-objective reinforce- ment learning (MORL) instead produces sets of policies that explicitly represent trade-offs between objectives. However, these policies are typically presented to the decision maker only through their value vectors, which can obscure substantial behavioral variation: poli- cies that induce distinct trajectories may appear indistinguishable when evaluated solely by expected returns. We propose an exploratory diagnostic workflow that automatically highlights behavioral variation along the Pareto front that objective values alone do not reveal, providing both quantitative and visual tools to support policy inspection. We vali- date our approach on simple grid examples and scale it to continuous control benchmarks, demonstrating that it remains effective as problem complexity increases.

BibTeX Entry

@article{Mone26TMLR_alignment,
    title={Objective-Behavior Alignment: Diagnostics for {MORL} Policy Selection},
    author={Antonio Mone and Zuzanna Osika and Florian Felten and Pradeep K. Murukannaiah and Mark Fuge and Frans A Oliehoek and Luciano Cavalcante Siebert},
    journal=    TMLR,
    issn={2835-8856},
    year={2026},
    OPTurl={https://openreview.net/forum?id=hfnMLNCCYz},
    keywords =  {refereed},
    note={},
    abstract = {
    Real-world decision-making often requires optimizing multiple competing objectives simul-
    taneously. In reinforcement learning (RL), this is typically addressed by combining reward
    signals into a single scalar objective via a scalarization function, which can be fragile: small
    changes in the weights can induce drastically different policies. Multi-objective reinforce-
    ment learning (MORL) instead produces sets of policies that explicitly represent trade-offs
    between objectives. However, these policies are typically presented to the decision maker
    only through their value vectors, which can obscure substantial behavioral variation: poli-
    cies that induce distinct trajectories may appear indistinguishable when evaluated solely
    by expected returns. We propose an exploratory diagnostic workflow that automatically
    highlights behavioral variation along the Pareto front that objective values alone do not
    reveal, providing both quantitative and visual tools to support policy inspection. We vali-
    date our approach on simple grid examples and scale it to continuous control benchmarks,
    demonstrating that it remains effective as problem complexity increases.
    }
}

Generated by bib2html.pl (written by Patrick Riley) on Thu Oct 08, 2026 13:51:42 UTC