Knowledge Transfer between Automated Planners

transfer problem-solving experience from previous tasks into the new task. Recently, the artificial intelligence community has attempted to model this transfer in an effort to improve learning on new tasks by using knowledge from related tasks. For example, classification and inference algorithms have been extended to support transfer of conceptual knowledge (for a survey see Torrey and Shavlik [2009]). Likewise, reinforcement learning has also been extended to support transfer (for a survey see Taylor and Stone [2009]; Torrey and Shavlik [2009]). This article presents an attempt to transfer structured knowledge in the framework of automated planning. Automated planning is the branch of artificial intelligence that studies the computational synthesis of ordered sets of actions that perform a given task (Ghallab, Nau, and Traverso 2004). A planner receives as input a collection of actions (that indicate how to modify the current state), a collection of goals to achieve, and a state. It then outputs a sequence of actions that achieve the goals from the initial state. Given that each action transforms the current state, planners may be viewed as searching for paths through the state space defined by the given actions. However, the search spaces can quickly become intractably large, such that the general problem of automated planning is PSpace-complete (Bylander 1994). The most common approach to coping with planning complexity involves defining heuristics that let the planner traverse the search space more efficiently. Current state-of-the-art planners use powerful domain-independent heuristics (Ghallab, Nau, and Traverso 2004; Nau 2007). These are not always sufficient, however, so an important research direction consists of defining manually or learning automatically domain-dependent heuristics (called control knowledge). In the latter case, Articles