Abstract
We describe a new forward search algorithm for classical planning. This algorithm attempts to maintain a focused search, expanding states using only a subset of the possible actions. Given a state s 0 that was obtained by applying action a to state s,
we prefer to apply in s 0 only actions a 0 that require some effect of a which we call forward actions. As this is incomplete, we must also consider actions a 00 that supply some other precondition of a 0 and actions a 000 that supply preconditions to a 00 and so on. We call these backward actions, as identifying the relevant actions requires backward reasoning. We show that by giving high priority to the forward actions a 0 we get improved performance in many domains. The resulting algorithm can be viewed as building on the classic idea of means-ends analysis [Newell and Simon, 1961]. One crucial open problem that arises is how to prioritize the search for backward actions.
we prefer to apply in s 0 only actions a 0 that require some effect of a which we call forward actions. As this is incomplete, we must also consider actions a 00 that supply some other precondition of a 0 and actions a 000 that supply preconditions to a 00 and so on. We call these backward actions, as identifying the relevant actions requires backward reasoning. We show that by giving high priority to the forward actions a 0 we get improved performance in many domains. The resulting algorithm can be viewed as building on the classic idea of means-ends analysis [Newell and Simon, 1961]. One crucial open problem that arises is how to prioritize the search for backward actions.
Original language | English |
---|---|
Pages | 44 |
Number of pages | 1 |
State | Published - 2017 |
Event | Heuristics and Search for Domain-independent Planning (HSDIP) - Pittsburgh, United States Duration: 20 Jun 2017 → 20 Jun 2017 |
Workshop
Workshop | Heuristics and Search for Domain-independent Planning (HSDIP) |
---|---|
Country/Territory | United States |
City | Pittsburgh |
Period | 20/06/17 → 20/06/17 |