Reinforcement learning (RL) is widely used to train decision-making agents in domains such as autonomous driving, healthcare, and game playing. However, RL agents typically optimize numerical rewards without explicitly accounting for ethical, social, or legal norms, which is problematic in safety-critical and dynamic settings. Policy fixing addresses this issue by augmenting RL agents with symbolic reasoning for detecting or anticipating norm violations and applying emergency plans minimally altering the learned policy. Answer Set Programming (ASP) enables fixes that balance reward optimization with norm satisfaction. Existing approaches, however, rely on batch reasoning, solving each step from scratch and struggling with temporal norms requiring information over time. In this paper, we propose an ASP-based Stream Reasoning extension supporting incremental, stateful processing of observations. The approach improves efficiency and naturally supports dynamic, time-bound normative obligations. Experimental results show improved scalability in demanding scenarios and effective temporal norm management.
Emergency Planning for Norm Violations via ASP-Based Stream Reasoning
Calimeri F.;Laboccetta L.;Perri S.;Terracina G.
2027-01-01
Abstract
Reinforcement learning (RL) is widely used to train decision-making agents in domains such as autonomous driving, healthcare, and game playing. However, RL agents typically optimize numerical rewards without explicitly accounting for ethical, social, or legal norms, which is problematic in safety-critical and dynamic settings. Policy fixing addresses this issue by augmenting RL agents with symbolic reasoning for detecting or anticipating norm violations and applying emergency plans minimally altering the learned policy. Answer Set Programming (ASP) enables fixes that balance reward optimization with norm satisfaction. Existing approaches, however, rely on batch reasoning, solving each step from scratch and struggling with temporal norms requiring information over time. In this paper, we propose an ASP-based Stream Reasoning extension supporting incremental, stateful processing of observations. The approach improves efficiency and naturally supports dynamic, time-bound normative obligations. Experimental results show improved scalability in demanding scenarios and effective temporal norm management.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.


