Loading...

Improving the Efficiency of Domain Independent AI Planning through Automatic Domain Knowledge Extraction

Akramifar, Ali | 2010

593 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 40700 (19)
  4. University: Sharif University of Technology
  5. Department: Computer Engineering
  6. Advisor(s): Ghassem-Sani, Gholamreza
  7. Abstract:
  8. Planning is an important branch of Artificial Intelligence. It is the area of study concerned with the automatic generation of plans to solve problems within a particular domain. At its simplest case, a plan is a total ordered sequence of actions. Given an initial state, planner tries to find the actions required to achieve some desired goal conditions. Planning technology has been successfully used in a variety of applications, including NASA's Remote-Agent Planner/Scheduler, which was practically employed in the Deep Space 1 spacecraft, satellite planning and scheduling for the European Meteostat, planning in a forest fire simulation system, etc. AI planning encompasses some difficulties such as inefficiency in solving real world problems, stemmed in the complexity of real world domains. The efficiency of an AI planning system can be improved by using some domain's knowledge. In some domain dependent planning systems, this type of knowledge is provided in advance by the user, and as a number of control rules to prune the search space. However, this is not acceptable in domain independent planning, in which any control knowledge should be learned automatically. There are two different approaches for applying learning techniques to AI planning: "planning learning" and "plan learning". In a plan-learning approach, some abstracted parts of previously produced plans are extracted and used in solving subsequent planning problems. On the other hand, in a planning-learning approach, some weights or rules are extracted and employed to make better decisions during future planning efforts. The focus of this thesis has been on the planning-learning approach to improve the efficiency of domain independent planning. It has been shown that there are some useful features, contained in the domain actions, which can be automatically extracted and used to improve the efficiency. To this purpose, three research phases were defined. First, to understand the behaviors and impacts of domain actions in a domain independent planning system, a domain independent planning system called HSP, which uses WA* as its search algorithm, was carefully studied. Then, an action related knowledge extraction method, named "voting" was designed and tested in a small domain specific planning system. Experimental results show some improvement in the efficiency of the planner. Finally, a new action related knowledge extraction method name least-failed-first heuristic was developed. This method was applied to the so-called enforced hill climbing search algorithm, which has been used to tackle problems from various standard planning domains. Our results show a significant improvement in the efficiency of the new planning method in many different domains
  9. Keywords:
  10. Machine Learning ; Artificial Intelligence Planning ; Increasing Efficiency ; Knowledge Extraction ; Enforced Hill Climbing ; Guided Enforced Hill Climbing

 Digital Object List

 Bookmark

No TOC