Applying Generalized Planning to Non-Markovian Problems: A Case Study of the Justified Perspectives Model in Epistemic Planning

Generalized Planning (GP) provides a scalable and reusable approach for addressing multiple classical planning instances within a specific domain. Unlike traditional methods that focus on individual instances, GP identifies common patterns across related scenarios to devise algorithmic solutions. However, GP faces significant challenges, especially in non-Markovian problems characterized by uncertainty and partial observability. In this context, Epistemic planning plays a critical role by incorporating knowledge and beliefs into decision-making, essential for managing non-Markovian complexities. This paper investigates the integration of the Justified Perspectives (JP) Model into GP to improve decision-making under non-Markovian conditions. The JP Model allows agents to develop and sustain justified beliefs from prior observations, which supports nested beliefs and the evaluation of epistemic goals. By combining JP with GP, the proposed method addresses both ontic and epistemic states, enabling more effective planning in non-Markovian environments. We evaluate the efficacy of this integrated approach in the coin domain by determining the minimal training set required and examining the influence of the training set sequence on planning outcomes. This integration significantly advances automated planning for real-world, uncertain scenarios, bridging the gap between theoretical models and practical applications in AI decision-making.