From: Stanley Kok (koks_at_cs.washington.edu)
Date: Tue Apr 15 2003 - 10:17:09 PDT
Paper Title: Refinement Planning as a Unifying Framework for Plan
Synthesis
Authors: Subbarao Kambhampati
One-line summary:
The paper presents refinement planning as a paradigm that abstracts
planning algorithms; summarizes four refinement strategies; and
provides an analysis of trade-offs in refinement planning.
Most Important Ideas in the Paper:
1. By classifying the refinement strategies, the papers allow us to
better appreciate the different strengths among them.
2. The identification of disjunctive representation as the major
strength of Graphplan reveals the importance of disjunctiveness
in improving performance.
Flaw in the Paper:
1. The examples are too simplistic to sufficiently illustrate some of
the ideas. For example, in Figure 20, it is not clear how promotion
and demotion can be applied to a plan with a single action.
Important, open research questions:
1. The use of learning techniques in customizing a planner's
behavior towards a problem population. It is not immediately
clear whether unsupervised learning can be used for this. In the
case of supervised learning, it will be interesting to see how
sufficient training samples can be obtained for learning to occur.
2. How refinement strategies can be extended for planning under
uncertainty.
This archive was generated by hypermail 2.1.6 : Tue Apr 15 2003 - 10:17:10 PDT