Instructors
Slides
Goals of the tutorial
-
Current planning algorithms can achieve impressive performance in many domains
and problems. However, there is still place for improvement among several
dimensions, such as consideration of quality, dynamic changes in domain and
quality descriptions, or integration of planning and scheduling (time and
resources reasoning). One can try to accommodate new requirements into
existing fast planning algorithms. Alternatively, the developer can
laboriously handcode the planning domain-dependent knowledge into some kind of
domain or control knowledge. In another perspective, machine learning can be
integrated with planning to automatically improve the planning performance
with experience.
In this tutorial we will: overview several planning algorithms in relation to
their learning opportunities; describe different structured ways of
representing learned knowledge; and present methods to combine planning and
learning to acquire domain and strategy knowledge, and to generate good
quality plans also considering time and resources. We will address both
classic deterministic, and nondeterministic planners.
Intended Audience
-
This tutorial is targeted at planning and scheduling researchers interested in
providing learning capabilities to their planning&scheduling systems. It is
also targeted at KR researchers given that we will address some knowledge
representation issues related to structures for representing knowledge, or
effects of representation changes with respect to learning and planning.
Detailed Contents
-
We will organize our tutorial as a four-hour intensive course on planning and
learning. We will support our presentation with online demonstrations of
planning and learning algorithms. We will illustrate the techniques with
realistic planning tasks, such as logistics/transportation, manufacturing
production planning, and information navigation domains.
- Motivation
- Goals of the tutorial
- Classic planning approaches
- Recent planning approaches
- Case-based planning and analogical reasoning
- Reactive planning
- Probabilistic planning
- Other approaches
- Domain descriptions: operators, macro-operators, policies, hierarchies of tasks
- Control knowledge: rules, policies, and temporal formulae
- Effects of representation changes in learning
- Domain-specific planners
- Explanation-based learning
- Static analysis of domain knowledge
- Plan reuse and derivational replay
- Inductive acquisition of control rules
- Genetic programming
- Probabilistic policy learning
- Hybrid techniques
- Evaluation metrics
- Incorporation of quality feedback
- Inductive refinement
- Observation and practice
- Execution and knowledge refinement
- Automatic learning of planners and policies from execution examples
- Review of big picture
- Summary of main points
[0:00 - 0:10] Complete Picture of the Tutorial
[0:10 - 0:30] Planning Algorithms and their Learning Opportunities
[0:30 - 0:50] Knowledge Representation in Learning for Planning
[0:50 - 1:50] Speedup Learning: Acquiring Control Knowledge for the Improvement of Planning Efficiency
[1:50 - 2:05] Break
[2:05 - 2:30] Learning Plan Quality
[2:30 - 3:45] Learning Domain Knowledge
[3:45 - 4:00] Wrap Up