## The Fast Downward Planning System (2006)

### Cached

### Download Links

Venue: | Journal of Artificial Intelligence Research |

Citations: | 230 - 27 self |

### BibTeX

@ARTICLE{Helmert06thefast,

author = {Malte Helmert},

title = {The Fast Downward Planning System},

journal = {Journal of Artificial Intelligence Research},

year = {2006},

volume = {26},

pages = {191--246}

}

### Years of Citing Articles

### OpenURL

### Abstract

Fast Downward is a classical planning system based on heuristic search. It can deal with general deterministic planning problems encoded in the propositional fragment of PDDL2.2, including advanced features like ADL conditions and effects and derived predicates (axioms). Like other well-known planners such as HSP and FF, Fast Downward is a progression planner, searching the space of world states of a planning task in the forward direction. However, unlike other PDDL planning systems, Fast Downward does not use the propositional PDDL representation of a planning task directly. Instead, the input is first translated into an alternative representation called multivalued planning tasks, which makes many of the implicit constraints of a propositional planning task explicit. Exploiting this alternative representation, Fast Downward uses hierarchical decompositions of planning tasks for computing its heuristic function, called the causal graph heuristic, which is very different from traditional HSP-like heuristics based on ignoring negative interactions of operators. In this article, we give a full account of Fast Downward’s approach to solving multi-valued planning tasks. We extend our earlier discussion of the causal graph heuristic to tasks involving

### Citations

11502 |
Computers and Intractability, A Guide to the Theory of NPCompleteness
- Garey, Johnson
- 1979
(Show Context)
Citation Context ...d also try to find sets of arcs of minimal total weight such that eliminating these arcs results in an acyclic graph. However, this is an NP-equivalent problem, even in the case of unweighted graphs (=-=Garey & Johnson, 1979-=-, problem GT8). After generating the pruned causal graph, we also prune the domain transition graphs by removing from the transition labels of DTG(v) all conditions on variables v ′ with v ≺ v ′ . The... |

4208 |
Artificial Intelligence: A Modern approach (2 nd Edition
- Russell, Norvig
- 2002
(Show Context)
Citation Context ...ponent of Fast Downward can perform its work in various alternative ways. There are three basic search algorithms to choose from: 1. Greedy best-first search: This is the standard textbook algorithm (=-=Russell & Norvig, 2003-=-), modified with a technique called deferred heuristic evaluation to mitigate the negative influence of wide branching. We have also extended the algorithm to deal with preferred operators, similar to... |

660 | The FF planning system: Fast plan generation through heuristic search
- Hoffmann, Nebel
- 2001
(Show Context)
Citation Context ...iscussion in Section 8. 2. Related Work As a planning system based on heuristic forward search, Fast Downward is clearly related to other heuristic planners such as HSP (Bonet & Geffner, 2001) or FF (=-=Hoffmann & Nebel, 2001-=-) on the architectural level. However, in this section we focus on work that is related on the conceptual level, i. e., work that uses similar forms of hierarchical decomposition of causal graphs and ... |

501 | 2.1: An Extension to PDDL for Expressing Temporal Planning Domains available at http://www.dur.ac.uk/d.p.long/pddl2.ps.gz
- Fox, Long
(Show Context)
Citation Context ... planner. Fast Downward is a classical planning system based on the ideas of heuristic forward search and hierarchical problem decomposition. It can deal with the full range of propositional PDDL2.2 (=-=Fox & Long, 2003-=-; Edelkamp & Hoffmann, 2004), i. e., in addition to STRIPS planning, it supports arbitrary formulae in operator preconditions and goal conditions, and it can deal with conditional and universally quan... |

470 | Planning in a Hierarchy of Abstraction Spaces - Sacerdoti - 1974 |

343 | Planning as heuristic search
- Bonet, Geffner
- 2001
(Show Context)
Citation Context ...esults in Section 7 and some discussion in Section 8. 2. Related Work As a planning system based on heuristic forward search, Fast Downward is clearly related to other heuristic planners such as HSP (=-=Bonet & Geffner, 2001-=-) or FF (Hoffmann & Nebel, 2001) on the architectural level. However, in this section we focus on work that is related on the conceptual level, i. e., work that uses similar forms of hierarchical deco... |

326 |
Linear-time algorithms for testing the satisfiability of propositional Horn formulae
- Dowling, Gallier
- 1984
(Show Context)
Citation Context ...ators associated with that branch. 5.3.2 AXIOM EVALUATORS Axiom evaluators are a simple data structure used for efficiently implementing the well-known marking algorithm for propositional Horn logic (=-=Dowling & Gallier, 1984-=-), extended and modified for the layered logic programs that correspond to the axioms of an MPT. They consist of two parts. Firstly, an indexing data structure maps a given variable/value pairing and ... |

307 | The computational complexity of propositional STRIPS planning
- BYLANDER
- 1994
(Show Context)
Citation Context ...rove that the restriction to unary operators and acyclic graphs does not reduce the complexity of plan existence: the problem is PSPACE-complete, just like unrestricted propositional STRIPS planning (=-=Bylander, 1994-=-). They also show that for singly connected causal graphs, shortest plans cannot be exponentially long, but the problem is still NP-complete. For an even more restricted class of causal graphs, namely... |

282 | GPS: A program that simulates human thought - Newell, Simon - 1963 |

210 |
Planning as search: A quantitative approach
- Korf
- 1987
(Show Context)
Citation Context ...perators are reversible and local search approaches like focused iterative-broadening search would be otherwise complete. In particular, search must fail in planning tasks which are not serializable (=-=Korf, 1987-=-). Therefore, if the first solution attempt fails, the algorithm is restarted without goal protection. The complete procedure is shown in Fig. 20, which concludes our discussion of Fast Downward’s sea... |

182 | Automatically generating abstractions for planning
- Knoblock
- 1994
(Show Context)
Citation Context ... are several transitions with identical conditions, we only keep one of them. 4. This simplification is closely related to Knoblock’s criterion for the problem-specific ordered monotonicity property (=-=Knoblock, 1994-=-). 23sHELMERT t1 t2 a1 a2 p1 Figure 13: Causal graph of a LOGISTICS task. State variables ti and ai encode the locations of trucks and airplanes, state variables pi the locations of packages. f1 f2 f3... |

155 | Complexity results for SAS+ planning
- Bäckström, Nebel
- 1995
(Show Context)
Citation Context ... is described in Section 6. 4. Multi-Valued Planning Tasks Let us now formally introduce the problem of planning with multi-valued state variables. Our formalism is based on the SAS + planning model (=-=Bäckström & Nebel, 1995-=-; Jonsson & Bäckström, 1998a), but extends it with axioms and conditional effects. Definition 1 Multi-valued planning tasks (MPTs) A multi-valued planning task (MPT) is given by a 5-tuple Π = 〈V,s0,s⋆... |

131 | Planning through stochastic local search and temporal action graphs in LPG
- Gerevini, Saetti, et al.
- 2003
(Show Context)
Citation Context ...wn excellent performance: The original implementation of the causal graph heuristic, plugged into a standard best-first search algorithm, outperformed the previous champions in that area, FF and LPG (=-=Gerevini, Saetti, & Serina, 2003-=-), on the set of STRIPS benchmarks from the first three international planning competitions (Helmert, 2004). Fast Downward itself followed in the footsteps of FF and LPG by winning the propositional, ... |

99 | A planning heuristic based on causal graph analysis
- Helmert
(Show Context)
Citation Context ...osition is not possible, because the subproblems that must be solved to achieve an operator precondition are not necessarily smaller than the original task. As far as we are aware, we were the first (=-=Helmert, 2004-=-) to present a general planning algorithm that focuses on exploiting hierarchical information from causal graphs. However, our causal graph heuristic also requires acyclicity; in the general case, it ... |

86 |
PDDL2.2: The language for the classical part of the 4th international planning competition
- Edelkamp, Hoffmann
(Show Context)
Citation Context ...nward is a classical planning system based on the ideas of heuristic forward search and hierarchical problem decomposition. It can deal with the full range of propositional PDDL2.2 (Fox & Long, 2003; =-=Edelkamp & Hoffmann, 2004-=-), i. e., in addition to STRIPS planning, it supports arbitrary formulae in operator preconditions and goal conditions, and it can deal with conditional and universally quantified effects and derived ... |

70 |
The HARPY Speech Recognition System
- Lowerre
- 1976
(Show Context)
Citation Context ...b-searches are again stopped, sub-searches for the third goal are started, and so on, until all goals have been reached. In some sense, our focusing technique is similar to the beam search algorithm (=-=Lowerre, 1976-=-), which also performs a fixed number of concurrent searches to avoid committing to a particular path in the search space too early. Beam search uses a heuristic function to evaluate which branches of... |

68 | Iterative broadening
- Ginsberg, Harvey
(Show Context)
Citation Context ...hieved goals in cond. This is an old idea called goal protection (Joslin & Roach, 1989). It is well-known that protecting 6. See the original analysis for a precise definition of “equally important” (=-=Ginsberg & Harvey, 1992-=-). While Ginsberg and Harvey’s assumption is certainly not valid in practice, we find it much more convincing than the competing model where goal states are uniformly distributed across the search fri... |

66 | The deterministic part of IPC-4: An overview
- Hoffmann, Edelkamp
- 2005
(Show Context)
Citation Context ...ur attention to the IPC4 planners. For this comparison, we chose the four most successful competition participants besides Fast Downward, namely LPG-TD, SGPlan, Macro-FF and YAHSP (cf. the results in =-=Hoffmann & Edelkamp, 2005-=-). Similar to the previous two experiments, we report the number of unsolved tasks in each domain (Fig. 29) and present graphs showing how quickly the tasks are solved (Figs. 30 and 31). Fast Downward... |

66 | A Reactive Planner for a Model-based Executive - Williams, Nayak - 1997 |

58 | Local search topology in planning benchmarks: A theoretical analysis
- Hoffmann
- 2002
(Show Context)
Citation Context ...eferred operators. There is a wealth of work on the FF heuristic in the literature, so we do not discuss it further. For a more thorough treatment, we point to the references (Hoffmann & Nebel, 2001; =-=Hoffmann, 2001-=-, 2002, 2005). 6.3 Greedy Best-First Search in Fast Downward Fast Downward uses greedy best-first search with a closed list as its default search algorithm. We assume that the reader is familiar with ... |

56 | Downward refinement and the efficiency of hierarchical problem solving. AIJ 71(1):43–100 - Bacchus, Yang - 1994 |

53 | Exhibiting knowledge in planning problems to minimize state encoding length
- Edelkamp, Helmert
- 1999
(Show Context)
Citation Context ...nning systems (van den Briel, Vossen, & Kambhampati, 2005). To keep the article focused, we do not discuss the translation algorithm here, referring to our earlier work for some of its central ideas (=-=Edelkamp & Helmert, 1999-=-). Instead, we consider its output, a multi-valued planning task, as a base formalism. • Second, no matter how clever the encoding is, most planning tasks are not completely hierarchical in nature. To... |

51 | Abstraction in Planning - Tenenberg - 1991 |

45 | Where ”ignoring delete lists” works: Local search topology in planning benchmarks
- Hoffmann
(Show Context)
Citation Context ... observe bad performance, we try to give a plausible explanation for this, but we do not conduct a full-blown study of heuristic quality in the spirit of Hoffmann’s work on the FF and h + heuristics (=-=Hoffmann, 2005-=-). While we do believe that much could be learned from such an investigation, it is a major undertaking that would go beyond the scope of this article. Our aim in this section is to evaluate the Fast ... |

42 | Structure and complexity of planning with unary operators
- Brafman, Domshlak
- 2003
(Show Context)
Citation Context ...hlak and Brafman provide a detailed account of the complexity of finding plans in the propositional STRIPS (with negation) formalism with unary operators and acyclic graphs (Domshlak & Brafman, 2002; =-=Brafman & Domshlak, 2003-=-). 3 Among other results, they prove that the restriction to unary operators and acyclic graphs does not reduce the complexity of plan existence: the problem is PSPACE-complete, just like unrestricted... |

38 | State-variable planning under structural restrictions: Algorithms and complexity
- Jonsson, Bäckström
- 1998
(Show Context)
Citation Context ...6. 4. Multi-Valued Planning Tasks Let us now formally introduce the problem of planning with multi-valued state variables. Our formalism is based on the SAS + planning model (Bäckström & Nebel, 1995; =-=Jonsson & Bäckström, 1998-=-a), but extends it with axioms and conditional effects. Definition 1 Multi-valued planning tasks (MPTs) A multi-valued planning task (MPT) is given by a 5-tuple Π = 〈V,s0,s⋆, A, O〉 with the following ... |

20 |
Multi-agent off-line coordination: Structure and complexity
- Domshlak, Dinitz
- 2001
(Show Context)
Citation Context ...y the extended graph shows how to change state from “freezing” (⊤) to “not freezing” (⊥). The example shows how plan execution can be viewed as simultaneous traversal of domain transition graphs (cf. =-=Domshlak & Dinitz, 2001-=-). This is an important notion for Fast Downward because the causal graph heuristic computes its heuristic estimates by solving subproblems of the planning task by looking for paths in domain transiti... |

20 | Reviving integer programming approaches for AI planning - Briel, Vossen, et al. - 2005 |

19 | Tractable plan existence does not imply tractable plan generation
- Jonsson, Bäckström
- 1998
(Show Context)
Citation Context ...6. 4. Multi-Valued Planning Tasks Let us now formally introduce the problem of planning with multi-valued state variables. Our formalism is based on the SAS + planning model (Bäckström & Nebel, 1995; =-=Jonsson & Bäckström, 1998-=-a), but extends it with axioms and conditional effects. Definition 1 Multi-valued planning tasks (MPTs) A multi-valued planning task (MPT) is given by a 5-tuple Π = 〈V,s0,s⋆, A, O〉 with the following ... |

18 | Incremental Planning - Jonsson, Bäckström - 1995 |

14 |
A Theoretical Analysis of Conjunctive-Goal Problems
- Joslin, Roach
(Show Context)
Citation Context ...ate reach-one-goal not to needlessly wander away from satisfied goals, we forbid applying operators that undo any of the previously achieved goals in cond. This is an old idea called goal protection (=-=Joslin & Roach, 1989-=-). It is well-known that protecting 6. See the original analysis for a precise definition of “equally important” (Ginsberg & Harvey, 1992). While Ginsberg and Harvey’s assumption is certainly not vali... |