Scholarlink

Acquiring computational specifications of human values is key for building value-aligned and value-aware AI systems. We surveyed methods for learning ethical principles and value-aligned behaviour from human demonstrations or specifications of values. However, to our knowledge, no attempt has been proposed for learning these value specifications from demonstrations, while satisfying the possibly diverse value preferences of different agents. In this work we propose a novel value learning framework (agnostic of specific learning techniques) for (i) learning value groundings (specifications of given values) and (ii) identifying value systems (the particular value preferences of agents) from observed behavior in an application domain. We illustrate our framework in a route choice modeling scenario, using tailored inverse reinforcement learning algorithms. The results show that we can successfully learn value systems that are coherent with observed route choices. Our findings open up intriguing concerns and challenges for future research in the area.

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

Value Learning for Value-Aligned Route Choice Modeling via Inverse Reinforcement Learning

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Andrés Holgado-Sánchez,
  • Javier Bajo,
  • Holger Billhardt,
  • Sascha Ossowski,
  • Joaquín Arias

摘要

Acquiring computational specifications of human values is key for building value-aligned and value-aware AI systems. We surveyed methods for learning ethical principles and value-aligned behaviour from human demonstrations or specifications of values. However, to our knowledge, no attempt has been proposed for learning these value specifications from demonstrations, while satisfying the possibly diverse value preferences of different agents. In this work we propose a novel value learning framework (agnostic of specific learning techniques) for (i) learning value groundings (specifications of given values) and (ii) identifying value systems (the particular value preferences of agents) from observed behavior in an application domain. We illustrate our framework in a route choice modeling scenario, using tailored inverse reinforcement learning algorithms. The results show that we can successfully learn value systems that are coherent with observed route choices. Our findings open up intriguing concerns and challenges for future research in the area.