Conclusions

• Planning graph can be seen as a CSP problem

  • The Dynamic CSP model corresponds closely to Graphplan’s backward search

• Adding EBL/DDB strategies to Graphplan’s backward search result in impressive speedups

  • EBL enables Graphplan to learn smaller and more useful memos
  • EBL/DDB capabilities are more useful than FC/DVO capabilities for Graphplan
  • Memos are a very conservative form of no-good learning
    • Sticky values help offset some disadvantages of this conservatism
  • EBL/DDB can also help in the context of random-restart search

Previous slide

Back to first slide

View graphic version