PhD Thesis - Flavio Tonidandel - 2002

 

Abstract

 

In this work a heuristic search case-based planning system, called Far-Off (Fast and Accurate Retrieval on Fast Forward), is presented. This system uses stored previous plans and the FF heuristic search planning system to find a solution to a new problem. For defining the three phases of the Far-Off system - Retrieval, adaptation and storing - several methods are developed and a new case-base maintenance policy is proposed. This policy, called minimal-injury, let the system to choose which cases can be inserted into or deleted from the case-base in order to keep its quality.

 

With many cases stored and structured in footprint cases and RelatedSets, it is possible to use the Footprint-based Retrieval (FbR) that decrease the space of searching for a similar case in a case-base.

 

For determining the similarity of a case to a new problem, a new similarity rule, called ADG (Action Distance-Guided), is designed. This rule is more accurate than the common case-based planners' similarity rules.

 

After the retrieval phase, the Far-Off system completes the retrieved case in order to turn it a potential solution to a new problem. Then, it applies the SQUIRE (Solution Quality Improvement by Replanning) in order to try to increase the quality of the solution.

 

 

 

FAR-OFF paper

- TONIDANDEL, Flavio; RILLO, Márcio. The FAR-OFF system: A heuristic search case-based planning. In:

                      INTERNATIONAL CONFERENCE ON AI PLANNING & SCHEDULING (AIPS), 2002, Toulouse - França.

                      Proceedings of AIPS 2002. AAAI Press, 2002  [pdf].

 


PUBLICATIONS related to FAR-OFF project

 

ADG method:

    

- TONIDANDEL, Flavio; RILLO, Márcio. An Accurate Adaptation-Guided Similarity Metric for Case-Based

                      Planning. Lecture Notes in Computer Science, Germany, v. 2080, p. 531-545, 2001. [pdf]

             - TONIDANDEL, Flavio; RILLO, Marcio. On the Use of the ADG Similarity in Case-Based Planning

                      Systems. In: ENIA´03 - IV ENCONTRO NACIONAL DE INTELIGÊNCIA ARTIFICIAL, 2003, Campinas.

                      Anais do XXIII Congresso da Sociedade Brasileira de Computação. 2003. [pdf]

 

 

 Min-Injury method:               

- TONIDANDEL, Flavio; RILLO, Márcio. Releasing Memory Space through a Case-deletion Policy with a

                      Lower bound for Residual Competence. Lecture Notes in Computer Science, Germany, v. 2080, p.

                      546-560, 2001. [pdf]

                                   

 Case Base Seeding method:   

- TONIDANDEL, Flavio; RILLO, Marcio. A Case Base Seeding for Case-Based Planning Systems. Lecture

                      Notes in Computer Science, v. 3315, p. 104-113, 2004. [pdf]

 

SQUIRE METHOD:

- TONIDANDEL, Flavio; RILLO, Marcio. Case Adaptation by Segment Replanning for Case-Based Planning Systems.

        In International Conference on Case-Based Reasoning – ICCBR 05. 2005. [pdf]

             - TONIDANDEL, Flavio; RILLO, Marcio. Improving the Planning Solution Quality by Replanning. In: 6.

                      SIMPÓSIO BRASILEIRO DE AUTOMAÇÃO INTELIGENTE, 2003, Bauru. 2003. [pdf]

                        

 

More Publications of Flavio Tonidandel are in the Publication´s Web Page.

 


DownLOADABLE FILES

 

A complete ZIP file, with CaseBases, Relations, HELP, Executable File of Far-OFF system, problems and domains can be download here

 

 FAROFF.ZIP

 

 

 

System Requirements

·      Windows 98/2000, XP or VISTA environment

·      Pentium III 450MHz

·      ~256 Mbytes RAM

·      ~2 Gbytes HD Free Space

·      Video Resolution: 1024x768 (mandatory)

 

INFORMATION

This implementation of the Far-Off system does not use dynamic variables. Instead, it uses global variables with some limited number of elements. The Far-Off system does not work with more than:

 

Domain features:                           System features

20 actions                                         5000 cases

30 predicates                                   3000 grounded predicates

100 different types                          8500 grounded actions

10 parameters of actions               500 actions in a plan

     or predicates                               100 retrieved cases