LAMA and Fast Downward

LAMA is a classical planner that finds a plan quickly and then keeps improving it. It is built on Fast Downward, the planning system that underlies much of classical planning research, and it won the satisficing track at IPC 2008 and IPC 2011.

Authors

Silvia Richter, Matthias Westphal (LAMA); Malte Helmert (Fast Downward)

Institutions

At publication: Griffith University and NICTA; University of Freiburg

Kind of planning

Classical, with action costs

Source

fast-downward.org, github.com/aibasel/downward (GPL-3.0)

How it works

Fast Downward translates a PDDL task into a representation with multi-valued variables, then searches it with heuristics. Its original paper introduced the causal graph heuristic, along with preferred operators and deferred evaluation.

LAMA adds a heuristic built from landmarks: as its paper puts it, "propositional formulas that must be true in every solution of a planning task". It searches with this and a variant of the FF heuristic together, first greedily and then with weighted A* at falling weights, so better plans keep arriving until the search is stopped.

LAMA is no longer a separate program. The Fast Downward README describes today's code as a merger of the original Fast Downward, LAMA and FD-Tech, and LAMA is run through the planner's aliases, such as lama and lama-first. Its documentation warns that these are not exactly the 2011 competition version, since the planner has changed since. Our benchmarks use the Fast Downward 24.06.1 release; newer releases exist.

What it handles

Classical PDDL, including ADL features, derived predicates and action costs. It does not handle numeric fluents or durative actions. See the feature comparison.

In the competitions

• Fast (Diagonally) Downward, by Malte Helmert and Silvia Richter, took first prize in the satisficing propositional track of IPC 2004.
• LAMA, by Silvia Richter and Matthias Westphal, won the sequential satisficing track of IPC 2008.
• LAMA won the satisficing track again in 2011, as LAMA-2011.

Results in our benchmarks

LAMA 24.06.1: problems solved with a plan the VAL validator accepts, within 60 seconds and 4 GB of memory on one core. These are our own runs, not competition results, and a planner written for other goals or for a different kind of problem can score low here while being the right tool for its purpose. How we measured.

Papers

1. Silvia Richter, Matthias Westphal. The LAMA Planner: Guiding Cost-Based Anytime Planning with Landmarks. Journal of Artificial Intelligence Research 39, 127-177, 2010.
2. Malte Helmert. The Fast Downward Planning System. Journal of Artificial Intelligence Research 26, 191-246, 2006.