Finding Diverse High-Quality Plans for Hypothesis Generation

New applications that use AI planning to generate explanations and hypotheses have given rise to a new class of planning problems, requiring finding multiple alternative plans while minimizing the cost of those plans. Hypotheses or explanations about a system, such as a monitored network host that could be infected by malware, are generated as candidate plans given a planning problem definition describing the sequence of observations and a domain model capturing the possible state transitions for the modeled system, as well as the many-to-many correspondence between the states and the observations. The plans must minimize both the penalties for unexplained observations and the cost of state transitions. Additionally, among those candidate plans, a small number of the most diverse plans must be selected as representatives for further analysis. To this end, we have developed a planner that first efficiently solves the “top-k” costoptimal planning problem to find k best plans, followed by clustering to produce diverse plans as cluster representatives. Experiments set in hypothesis generation domains show that the top-k planning problem can be solved in time comparable to cost-optimal planning using Fast-Downward. We further empirically evaluate multiple clustering algorithms and similarity measures, and characterize the tradeoffs in choosing parameters and similarity measures.