From Theory to Practice: Foundations for Scalable and Human-Aligned Multi-Objective Reinforcement Learning

Reinforcement learning (RL) has demonstrated remarkable success in complex domains. Yet, its
industry adoption remains limited due to two critical obstacles: the difficulty of designing effective
reward schemes and the inability to adapt efficiently to modified or new tasks without extensive
retraining. This challenge is particularly important in real-world scenarios, which typically involve
multiple competing objectives. For instance, in water management systems, decision-makers have to
balance flood control, irrigation, and ecological preservation. While multi-objective reinforcement
learning (MORL) offers a more intuitive framework by learning optimal behaviours for different tradeoffs,
existing approaches remain computationally expensive, typically requiring multiple training runs
and costly re-training when objectives change.
This proposal bridges the gap between academic innovations and industry applications through three
key contributions: (1) a novel MORL framework enabling rapid adaptation to new or shifting
objectives without extensive retraining; (2) interactive methods helping decision-makers translate
their preferences into MORL frameworks; and (3) guiding principles for practical MORL application,
validated through a water management case study. By integrating these innovations, this research
will make MORL more sample-efficient, adaptive, and accessible, making it a viable tool for realworld
decision-making across industries.