Scholarlink

In many real-world problems, one faces the problem of having to make decisions considering several conflicting objectives. In such problems, the solution is a set of policies rather than a single one. This leads to multi-objective reinforcement learning problems (MORL) which have not received much attention until recent years by the reinforcement learning community. In this work, we propose coupling the R2 indicator with Pareto Q-learning. The R2 indicator has been successfully used for multi-objective optimization problems making it a good candidate for MORL. We tested our approach on several problems from MO Gymnasium and compared it with HB-MORL, which uses the hypervolume indicator. Our preliminary results show that the novel algorithm obtains competitive results and that could be an interesting alternative when dealing with MORL problems.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A R2 Based Multi-objective Reinforcement Learning Algorithm

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Sofia Magdalena Borrel Miller,
  • Carlos Ignacio Hernández Castellanos

摘要

In many real-world problems, one faces the problem of having to make decisions considering several conflicting objectives. In such problems, the solution is a set of policies rather than a single one. This leads to multi-objective reinforcement learning problems (MORL) which have not received much attention until recent years by the reinforcement learning community. In this work, we propose coupling the R2 indicator with Pareto Q-learning. The R2 indicator has been successfully used for multi-objective optimization problems making it a good candidate for MORL. We tested our approach on several problems from MO Gymnasium and compared it with HB-MORL, which uses the hypervolume indicator. Our preliminary results show that the novel algorithm obtains competitive results and that could be an interesting alternative when dealing with MORL problems.