Learning-driven goal generation.

Automated Planning deals with reasoning processes where a set of goals must be achieved from an initial state using some actions. Most work on planning have a static view of goals; they are given at start of the planning process and they do not change over planning and/or plan execution. However, in many real world domains, agents need to consider dynamic goal management. In this paper, we propose to increase the performance of planning agents by learning when goals will appear in the near future. The learned predictive models allow agents to perform some kind of anticipatory planning, where the planning process considers not only current goals, but also future predicted goals. We also study under which conditions this anticipatory approach outperforms a standard planning approach. Finally, experiments that support our hypothesis are presented.