BackgroundDESCARWIN is an ANR project coordinated by Pierre Savéant (Thalès). Beside TAO, it also involves Vincent Vidal at ONERA (Toulouse).
DESCARWIN - the hybridization of Descartes and Darwin - is based on Divide and Evolve, an original approach to Temporal Planning Problems (TPPs) solving (see the seminal paper here). The target TPP is decomposed into a sequence of (hopefully simpler) TPPs by means of artificial evolution, and each TPP of the series is handled by a standard solver.
Work PlanWithin DESCARWIN, the existing Divide and Evolve (DAE) framework will be extended in several directions.
- Domain knowledge gathered from the initial TPP should be used to improve all evolutionary operators, from initialization to crossover and mutation
- The parameters of the Evolutionary Algorithms need to be automatically adapted to the domain, or even to the instance. An important issue regards the identification of the characteristics of a domain that are common to all isntances - if any.
- Different planners can be used to solve the small TPPs, and should be compared within DAE. Ultimately, the evolutionary algorithm itself can be used to chose, for each sub-TPP, the most efficient planner.
- Using Evolutionary Algorithms opens the path to Multi-Objective optimization (e.g. the duration of a plan and its cost, or risk, or ...). Though the proof of concept has already been proposed in the seminal work, ot remains to be validated on large TPP. In particular, there is a need for a benchmark suite to be designed.
PrerequisitesThe candidate should preferably have a strong background in (temporal) AI Planning and Constraint Programming. Some knowledge about Evolutionary Computation would be ideal, but is not required. Experience and skills in C++ programming are mandatory.
- The position is available immediately (January 2010)
- The post-doc will be located at LRI in the INRIA project-team TAO
Contact: Marc Schoenauer