Lecture Notes in AI Planning

Lecture Notes

• Planning, Execution, and Learning taught by Manuela M. Veloso at CMU, Fall, 2008.
• AI Planning taught by Dana S. Nau at University of Maryland, Spring, 2007. This course use the Automated Planning: theory and practice as textbook.
  • Classical Planning
  • Complexity of Classical Planning
  • State-Space Planning
  • Plan-Space Planning
  • Planning-Graph Techniques
• AI Planning in AI course by Alison Cawasey at Heriot Watt University, 2007.
• AI Planning given by Mathijs de Weerdt at Delt University of Technology, 2005.
• AI Planning course taught by Meirav Hadad at Haifa University, 2005.
• AI Planning taught by Jim BlytheJose-Luis Ambite, and Yolanda Gil at University of Southern California’s Information Sciences Institute, 2003.
  • Plan generation, Plan space, and causal-link planning, UCPOP
  • GraphPlan
• AI lecture notes by Bruce Randall Donald at Dartmouth Computer Science , 1998.
  • Planning study guide: an introduction of planning systems, types of planner, like as logic planners, non-hierarchical/hierarchical non-linear/linear planners. Issues and timeline of planners have been given in this literature.
  • UCPOP background: a summary of the Daniel Weld article.
• Cognitive Oriented AI
  • Basic AI planning: STRIPS, state-space model, planning algorithms (forward/backward (or regression) planning)
  • Planning-graph techniques: Deductive planning, situation calculus, GraphPlan

AI Planner Collections by Robert St. Amant in North Carolina State University, this site has been moved to AIwiki site.

Graphplan Open Sources in Java

Graphplan is a general-purpose planner for STRIPS-style domains, based on ideas used in graph algorithms. This alogirhtm was first developed by Avrim Blum and Merrick using C in 1995. There are some researchers implement Graphplan in Java:

1. PL-PLAN can be used as Java library, implemented by Philippe Fournier-Viger and Ludovic Lebel in 2004 summer. The .jar file contains source codes and test data without any Java class or library, user should compile and build this project first. In my point of view, their implementation have the following problems:
   • No parameter passing support: PL-PLAN does not provide parameter list for operator definition. Users should declare different actions for matching the objects in the parameter list, but it’s hard to enumerate all the possible combinations when there are many operators and objects. For example, if a problem contains single type of object with 4 instances and there is an operation takes 2 parameters, user should combine two different instances to associate with this operation’s parameters and it needs 4*3=12 times.
   • Unexecutable testing program: Some JUnit test codes encountered ‘undefined method’ compile errors.
     
   • Poor file organization: Files are not organized well, either testing data or some Java programs are stored in wrong folder. For example, the demo programs Example1.java and Example2.java are located in javaapi directory, these sample programs should store in a specific demonstration folder. The Graphplan API (javaapi) directory definitely not the right folder for such example program files. Moreover, the testing data files are allocated on the PL-PLAN root directory, these files must be stored in a specified directory as the data repository. An unit test folder should be prepared for the JUnit test programs instead of storing in API (or algorithms) folder .
2. JPLAN can be downloaded from SourceForge.net, implemented by By Yasser EL-Manzalawy in April, 2004. The .jar code consists of a runnable library and source codes. Simply run one of the three example programs, rocket, blocks, and robot, will get .gp and .pln files. The .gp is the mutual exclusive description of the leveled GraphPlan graph, whereas the .pln is the planning results. The following is my summary of JPLAN:
   • Support parameter matching, but not follow the PDDL (Planning Domain Definition Language)  standard.
   • File organization is better than PL-PLAN
   • All programs can be successfully compiled
   • The token parsing programs contains goto-like statements, programs contain <label> and continue <label> statements for parsing input problem.
   • The naming and programming style of parser programs are not as good as the styles of Graphplan algorithm programs. The author uses meaningless word jj as the variable and method names.