Cost-Based Learning for Planning

Most learning in planners to date has been focused on speedup learning. Recently the focus has been more on learning to improve plan quality. We introduce a different dimension: learning not just from failed plans, but learning from inefficient plans. We call this cost-based learning (CAL). CBL can be used to improve both plan quality and provide speedup learning. We show how cost-based learning can also be used to learn plan rewrite rules that can be used to rewrite an inefficient plan to an efficient one, in the style of Planning by Rewriting (PbR). We do this by making use of dominance relations. Additionally, the learned rules are compact and do not rely on state information so they are fast to match.