Robust Multiobjective Reinforcement Learning Considering Environmental Uncertainties.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Optimization

Background:

• Real-world problems often involve multiple conflicting objectives requiring preference weighting.
• Environmental uncertainties like changes or noise can lead to suboptimal policies despite aiming for Pareto optimality.

Purpose of the Study:

• To present a novel robust multiobjective reinforcement learning (RMORL) paradigm.
• To train a single model capable of approximating robust Pareto-optimal policies across diverse preference spaces.

Main Methods:

• Modeled environmental disturbance as an adversarial agent within a zero-sum game integrated into a multiobjective Markov decision process (MOMDP).
• Developed an adversarial defense technique against observational perturbations to bound policy variations under specified preferences.

Main Results:

• The proposed RMORL technique was evaluated in five multiobjective environments with continuous action spaces.
• Demonstrated effectiveness through comparisons against classical and state-of-the-art baseline methods.

Conclusions:

• RMORL effectively enhances policy robustness against environmental uncertainties and observational perturbations.
• The approach enables a single model to achieve robust Pareto optimality across the preference space.