[2Yin-A] Poster 3

[2Yin-A-39] Reinforcement learning agent considering intrinsic value in obligation constraints

〇Yotaro Nakayama¹, Yoshitaka Aoki¹, Seiki Akama² (1. BIPROGY Inc., 2. C-Republic Inc.)

Keywords:

autonomous agent,deontic logic,intrinsic value,reinforcement learning

In recent years, research into machine ethics has expanded, focusing on how ethical behavior can be achieved in autonomous AI. In this study, we propose a theoretical framework for intrinsic reward expected act utilitarianism, in which agents internally maintain their own value assessments, and obligation constraints are reflected as subjective rewards. By incorporating the possibility of violating obligations into their own reward structure and updating their self-aware reward predictions, agents learn to balance compliance with obligations and goal achievement. We conduct reinforcement learning experiments that introduce subjective rewards based on intrinsic value in addition to extrinsic rewards, and verify the impact that differences in intrinsic value have on policy formation and exploration behavior under obligation constraints. By considering agents' intrinsic value rewards in interpreting obligation constraints, decision-making can account for the subjectivity of rewards in both behavior and the environment.