Back to the HomePage.

Who uses Fast Downward?

The following list shows some of the uses that people have found for the Fast Downward and LAMA planners.

We have only just started collecting this list, so as of now it is very incomplete.

• Blocks World game for the iPad: [2012]

• Minh Do and Minh Tran (of TranCreative Software) used Fast Downward to generate base scores for all levels of the Blocks World iOS game. This game involves a variation of Blocks World with limited number of table spaces and possibly duplicate letter blocks. Different levels/problems represent different anagrams (e.g., "THE EYES = THEY SEE"). The base scores calculated by Fast Downward are used to rate players performance in the same way base scores are used to rate planners in the IPC. Being able to give players sensible scores to grade their performances is a very important aspect of gameplay.

• References: Facebook page and iTunes page of the game

• proving properties of cellular automata: [2010]

• Jörg Hoffmann, Nazim Fatès and Héctor Palacios encode the problem of finding fixed point configurations for certain asynchronously updating cellular automata as a classical planning problem. They use different planning systems, including LAMA, to find such fixed points. In their experiments, LAMA performs better than FF but worse than SATPLAN, indicating that this a domain where current heuristic search approaches fare atypically badly.

• Reference: Hoffmann, J., Fatès, N., & Palacios, H. (2010). Brothers in Arms? On AI Planning and Cellular Automata. In Proc. ECAI 2010, pp. 223-228.

• generating plans for greenhouse logistics [2010]

• Malte Helmert and Hauke Lasinger describe a planning problem that arises when planning the travel routes for plants in smart greenhouses. They compare the handcrafted system currently used for generating plans for this application to domain-independent planners, including LAMA. Their results show that LAMA scales quite well compared to the handcrafted system and even outperforms it in some cases, while also being flexible enough to adapt to more complex application requirements.

• Reference: Helmert, M.; & Lasinger, H. (2010). The Scanalyzer Domain: Greenhouse Logistics as a Planning Problem. In Proc. ICAPS 2010, pp. 234-237.

• natural language sentence generation [2010]

• Alexander Koller and Jörg Hoffmann describe how to express the problem of natural-language sentence generation as a classical planning task and use FF, LAMA and a modified version of FF to solve the resulting planning tasks. In their experiments, LAMA outperforms the unmodified version of FF by several orders of magnitude, but is in turn outperformed by the modified version of FF by several orders of magnitude.

• Reference: Koller, A., & Hoffmann, J. (2010). Waking Up a Sleeping Rabbit: On Natural-Language Sentence Generation with FF. In Proc. ICAPS 2010, pp. 238-241.

• over-subscription planning (planning with soft goals) [2009]

• Emil Keyder and Héctor Geffner describe a technique for compiling away soft goals (à la the IPC-2008 net benefit track) and apply LAMA to the resulting problems. The LAMA-based planner performed drastically better than native over-subscription planners.

• Reference: Keyder, E., & Geffner, H. (2009). Soft Goals Can Be Compiled Away. JAIR, 36, 547-556

• finite-state controller synthesis [2009]

• Blai Bonet, Héctor Palacios and Héctor Geffner show how to solve control problems by compiling the problem of generating a memoryless or finite-state controller into a classical planning task. These classical planning tasks are then solved with LAMA or SATPLAN. (In their experiments, neither planner dominates the other.)

• Reference: Bonet, B., Palacios, H., & Geffner, H. (2009). Automatic Derivation of Memoryless Policies and Finite-State Controllers Using Classical Planners. In Proc. ICAPS 2009, pp. 34-41.

• massively parallel planning [2009]

• Akihiro Kishimoto, Alex Fukunaga and Adi Botea have parallelized Fast Downward and run the distributed version on a SunFire X4600 cluster consisting of 8 machines with a total of 64 CPUs, 128 CPU cores and 256 GB RAM. They typically achieve a search speedup of 30-60 compared to a sequential version of the planner.

• Reference: Kishimoto, A., Fukunaga, A., & Botea, A. (2009). Scalable, Parallel Best-First Search for Optimal Sequential Planning. In Proc. ICAPS 2009, pp. 201-208.