9:30 - 10:30: Task planning and agent architectures (Philippe Morignot) 10:30 - 11:00: Pumas project presentation (Jean-Marc Lasgouttes)
An agent often has to perform several successive actions in order to satisfy future goals. For example, if my goal is to be at home, I have to open the door of my office, walk to my car, hold the keys, open the driver’s door, sit down, close the door, start the engine and drive home. Task planning is a sub field of Artificial Intelligence dedicated to modeling this problem and finding algorithms which quickly build such action plans, as solution to planning problems.
In the first part of this talk, we present what the classical planning problem and its assumptions are, why complexity results make it difficult, and introduce the planning domain definition language (PDDL) dedicated to representing a planning problem. As an example of use of this language, a plan will be built step by step in the blocks world (stacking blocks on a table in order to reach some given configuration). An overview of existing algorithms for building a classical task planner will be presented: State-space search, plan-space search, SAT-based planning, CSP-based planning, hierarchical tasks network and the GRAPHPLAN planner. Finally, conditional planning will be presented.
The second part of this talk is dedicated to using plans in dynamic, unpredictable and real environment — and not building plans under classical assumptions, as this is the case in the first part. Building a plan, and then executing it (the “plan-then-execute” paradigm) is inappropriate in dynamic, unpredictable, real environments where things usually do not unfold as expected: Re-planning, at least, and interleaving planning and execution is required. We present several existing agent’s architectures, from reactive ones (e.g., Brooks’ subsumption architecture) to cognitive ones (e.g., 2-level, 3-level, ontology-based).