Researchers have introduced MO-GRPO, a novel algorithm designed to address reward hacking in multi-objective reinforcement learning. This new method extends the Group Relative Policy Optimization (GRPO) by incorporating a simple normalization technique to automatically reweight reward functions based on their value variances. MO-GRPO ensures that all objectives contribute equitably to the loss function, eliminating the need for manual scaling and preserving preference order. Experimental evaluations across various domains, including multi-armed bandits, simulated control tasks, machine translation, and instruction following, demonstrate MO-GRPO's ability to achieve stable learning and outperform standard GRPO. AI
IMPACT Introduces a method to improve stability and performance in multi-objective reinforcement learning tasks.
RANK_REASON This is a research paper detailing a new algorithm for reinforcement learning. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
