This paper focuses on learning action costs, rather than learning action dynamics, in behavioral model learning. Unlike previous studies that focused on specifying valid plans for planning tasks, this paper presents a novel problem: learning a set of action costs that ensures that a set of input plans is optimal under the resulting planning model. To address this problem, we propose $LACFIP^k$, an algorithm that learns action costs from unlabeled input plans. We demonstrate the successful performance of $LACFIP^k$ through theoretical and experimental results.
