An important data source for planning is the traces left by plan executions. It is an interesting and useful problem to mine high-utility plans from the plan traces for future planning applications, in areas such as marketing planning in customer relationship management . Traditional data mining algorithms focus on finding frequent sequences without utility consideration through sequential mining, whereas traditional Markov Decision Process based learning and planning systems are faced with high computational cost. In this paper, we present a novel algorithm, which automatically generates plans from large databases by combining data mining and AI planning. Our objective is to find high-utility plans that convert groups of records from less desirable states to more desirable ones. Our al...[ Read more ]

An important data source for planning is the traces left by plan executions. It is an interesting and useful problem to mine high-utility plans from the plan traces for future planning applications, in areas such as marketing planning in customer relationship management . Traditional data mining algorithms focus on finding frequent sequences without utility consideration through sequential mining, whereas traditional Markov Decision Process based learning and planning systems are faced with high computational cost. In this paper, we present a novel algorithm, which automatically generates plans from large databases by combining data mining and AI planning. Our objective is to find high-utility plans that convert groups of records from less desirable states to more desirable ones. Our algorithm focuses on an abstract space of the original problem during planning to enable efficient and effective approximation. Our formulation and solution avoid both the shortcomings of traditional AI planning, which relies on the precise knowledge of actions’ logical formulations, and the computational problem faced by many Markov decision process formulations in uncertainty planning. We show through empirical results that planning using our combined algorithm produces high-utility plans that offer a tradeoff between optimality and efficiency.