Techniques and Knowledge used for Adaptation during ... - CiteSeerX

Techniques and Knowledge used for Adaptation during Case-Based Problem Solving

T

Wolfgang Wilke and Ralph Bergmann

Keywords:

Topics:

AF

University of Kaiserslautern Centre for Learning Systems and Applications (LSA) Department of Computer Science P.O. Box 3049, D-67653 Kaiserslautern, Germany E-Mail: fwilke,[email protected]

Case-Based Reasoning, Adaptation, Knowledge Modelling

1.1 Knowledge Modelling, 6.2 Analogy and Case-Based learning, 6.3 Multistrategy and hybrid

formulations

DR

Preferred presentation: oral

Acknowledgements

The authors would like to thank: Padraig Cunningham, Harald Holz, Hector Mu~noz-Avila, Jurgen Schumacher, Barry Smith and Armin Stahl for their helpful discussions and comments on the topic of the paper. This work was partially funded by the Commission of the European Communities (ESPRIT contract P22196, the INRECA II project: Information and Knowledge Reengineering for Reasoning from Cases) with the partners: AcknoSoft (prime contractor, France), Daimler Benz (Germany), tecInno (Germany), Irish Medical Systems (Ireland) and the University of Kaiserslautern (Germany).

1 Introduction

AF

T

Case-based reasoning (CBR) has recently become a rising star in knowledge-based decision support technology [Thompson, 1997]. It has been used to create numerous applications in a wide range of domains, including prediction, diagnosis, planning, process and quality control, monitoring, supervision, technical maintenance and software systems for decision or customer support. This results from the viewpoint that CBR is seen as a cognitively sound modelling approach for explaining human problem solving in domains where experience plays an important role and as a software engineering approach for how to model and implement decision support systems that are able to use past experiences for suggesting solutions or making predictions. It is especially appealing to those professionals who solve problems by recalling what they did in similar situations that happened in the past. The main hypothesis behind CBR is quite simple: "similar problems have similar solutions" - or put it the other way round - "you can reuse the solution of a similar problem in order to solve your actual problem". Based on this hypothesis the adaptation of the solution from the most similar problems should be a very ecient and straightforward way to come to a suggested solution for the actual problem. But in reality it is not always so easy.

DR

While the theory and practice of case-based reasoning has bene ted greatly from recent advances in case representation, similarity assessment, and retrieval, adaptation is still considered as the most dicult step in CBR. The reason for this is that there is a large variety of dierent adaptation techniques, all having dierent requirements on available adaptation knowledge. While it is often easy to acquire the cases, the required adaptation knowledge is often very hard to get. Therefore, one strategy for building CBR applications is to bypass adaptation entirely [Leake, 1996] which can be appropriate for analytic tasks like classi cation, diagnosis, or decision support. Instead, one tries to blow-up the case base to ensure that for every problem a suciently similar case is available. However, for synthetic tasks (like con guration, design, or planning), the vast space of possible problems and solutions disables this approach and adaptation becomes essential. This paper presents an overview of dierent adaptation methods which are common in today's systems. We introduce rst the process model of CBR and the used knowledge according to the dierent knowledge containers. Next, we present current models of adaptation and illustrate them in an example domain and close with some remarks.

2 Knowledge Modelling for Case-Based Systems Before we have a closer look at dierent adaptation techniques we have to understand the CBR process during problem solving and the used knowledge sources. This is illustrated in Figure 1.

T

Richter [Richter, 1995] describes four containers in which a CBR system store knowledge. The four containers are: the vocabulary knowledge used to describe the whole domain, the cases knowledge of the cases in the case base, the retrieval knowledge used for the retrieval of similar cases and the adaptation knowledge used during the transformation of retrieved solutions. During problem solving, rst a query (Q) is entered by the user. Next, the most similar problem descriptions (S) are retrieved from the case base. To determine similar problems and their solutions retrieval knowledge, like the indexing of the cases or the similarity measure is used. The case knowledge is represented in the case base, which consists of all known problems and their belonging solutions. The system takes the correlating soluFig. 1. The CBR process and the used knowledge. tions from the past and adapts them to solve the new problem. Therefore, adaptation knowledge is applied that describes how to transform the similar solutions for solving the query problem. Often, adaptation is a necessity and makes problem solving still tractable or even possible at all. This results from the possibility to broaden the coverage of past cases by adaptation knowledge for new problems. Which and how knowledge is represented constraints the coverage of the whole CBR system and establishes the vocabulary knowledge. S

S

Q

Retrieval Knowledge

Problem Space

AF

Case Knowledge

Vocabulary Knowledge

S

Adaptation Knowledge

DR

Solution Space

3 Current Models of Adaptation

In this section we have a closer look at several adaptation techniques which are used in CBR. To demonstrate these techniques, we will rst introduce an example domain were personal computers (PC's) are con gured.

3.1 The Example Domain: PC-Con guration

The task of PC con guration is to arrange a personal computer for the special customer needs. A related CBR application can support shop assistants to sell optimal con gurations as well as possibly retail computers via the Internet. The customer speci es his requirements by denoting which applications he would like to use the PC for. For capturing this requirements, we introduce set of attributes for application areas like word-processing, database, music, programming or games.

T

Each of these attributes has a certain value denoting the importance of this application area for the user. The goal is to con gure a PC which can be used for the desired mixture of applications. The solution itself, i.e. the nal PC con guration, is represented using an object-oriented model of the PC and its components.

3.2 The Continuum of Adaptation Models

AF

From a case-based reasoning perspective there is a natural relationship between the complexity of the problem solving task and the complexity of the CBR adaptation process. Even adaptation is a special kind of problem solving for the desired task with the helpful input of a problem description, similar problems and a belonging similar solutions. The underlying hope is, that these useful input together with a special optimised problem solving algorithm for the given information shorten the problem solving process against problem solving from scratch. An overview of several basic

Adaptation

Null Adaptation

Generative Adaptation

Transfomational Adaptation

Substitutional Adaptation

Structural Adaptation

Initalization

One-Shot Replay

DR

Solution Trace Replay

Fig. 2.

Interleaved Replay

Complete Decision Replay

The continuum of adaptation models

adaptation approaches which are exploited in recent research is given in Figure 2. We observe an increasing complexity from null adaptation on the left, over transformational approaches to generative approaches. Further using multiple similar cases for adaptation is called compositional adaptation. We will now describe these dierent approaches and illustrate them with an example in our problem domain.

3.3 Null Adaptation

The simplest kind of adaptation is null adaptation. The case base is searched for a similar case containing the most similar user requirements and the solution from this case is taken to solve the

current problem without any modi cation. Null adaptation means: the adaptation is left to the user or it is even not necessary to adapt at all. Most commercial CBR systems work with null adaptation.

T

3.4 Transformational Adaptation

AF

Transformational adaptation means that the old solution of the similar case is transformed into a new solution for the new problem. Transformational adaptation [Carbonell, 1983] supports the reorganisation of solution elements and permits the modi cation, addition and deletion of these elements under certain conditions. Typically, transformational adaptation systems employ a xed set of adaptation operators and/or transformation rules. Depending on the dierence of problem attributes in the query and the problem attributes in the similar case, the actual solution is modi ed. Thus, transformational adaptation requires domain knowledge on how certain dierences in the problem lead to dierences in the solution. Depending on the degree of modi cation we distinguish between substitutional adaptation and structural adaptation. Transformational adaptation is for example used in: Clavier [Barletta and Hennessy, 1989, Hennessy and Hinkle, 1991] or INRECA [Wilke and Bergmann, 1996].

Substitutional Adaptation For CBR systems concerned with simple problem solving tasks simple

DR

substitutional adaptation is sucient. It is promising when the retrieved case will typically be very close to the target and consequently will require only minor modi cations. In this adaptation model it is only possible to change the values of attributes. The structure of the new solution remains unchanged. Of course, the values of the attributes are reused from the old case, too, but it is possible to change some of them to get a solution of a better quality. Here adaptation means a recalculation of several parameters depending on the relation of the attributes of the problem description of the query and the similar case. In our example domain, a possible substitutional adaptation is the modi cation of the hard disk space depending on the usage of database systems with the con gured PC. If the value for database usage in the problem descriptions diers there is the necessity to add hard disk space. This could be coded into simple rules, like:

query:database similar:database > 5 ) solution:diskspace = solution:diskspace + 2GB

Detailed descriptions of substitutional adaptation can be found in [Bain, 1986].

Structural Adaptation With routine problem solving tasks, the retrieved case, will probably re-

quire more substantial modi cations. If we allow to change the structure of the solution during adaptation we speak of structural adaptation. Structural adaptation [Goel and Chandrasekaran, 1989,

AF

T

Smyth and Keane, 1996] supports the reorganisation of solution elements and permits the addition and deletion of such elements under certain conditions. Also structural adaptation systems employ a xed set of adaptation operators and/or transformation rules which modify the structure of the solution, depending on relations between the problem description of the query and the similar case. Structural changes can occur when complete components are added or deleted or when another object replace an existing one. An example: The similar case being retrieved was not intended for playing computer games. One of the components which is needed only for games is a joystick. Therefore the solution in the old case does not contain a joystick. If the customer demands computer games in the query, during adaptation an instance of the object joystick must be added and con gured to make the old case suitable for the new query. An adaptation rule which realises such a structural modi cation would be:

query:games > 0 and similar:games = 0 ) AddObject : Joystick

In contrast to the rst example, joystick is a new complex object in the object-oriented model of the solution and not an already existing part of the solution which is modi ed only.

3.5 Generative Adaptation

DR

Generative adaptation is radically dierent from transformational adaptation. Generative adaptation requires a generative from-scratch problem solver that is tightly integrated into the case-based reasoner for the purpose of adaptation. Typically, such a problem solver (e.g., a con guration system in the introduced PC domain) is in principle able to solve the type of problem that should be handled by the case-based system alone, i.e., without the use of cases. However, in practice such a pure generative problem solver is usually insucient because of the computational complexity of the generative problem solving process or because of the insucient quality of the solutions it produces. During generative adaptation, the generative problem solver is of course not used to solve the whole new problems from scratch, but only to generate those parts of the solution that are inadequate due to dierences between the current problem and the retrieved case. Thus, drawbacks like lack of eciency or quality don't have a great impact as long as the retrieved case is suciently similar to the current problem. As a consequence of the use of a generative problem solver, generative adaptation requires a dierent kind of knowledge than transformational adaptation. Instead of having knowledge that describes how dierences in the problem description lead to dierences in the solution, knowledge that allows to construct a solution from scratch is required. The particular knowledge needed is strictly dependent from the generative problem solver that is used. For example, in the introduced

AF

T

PC domain, a con guration systems needs knowledge about the components like the function they ful l and the constraints that exist between dierent components. For a particular graphics adapter, for instance, the following knowledge must be available: The Matrox Millenium graphics adapter is suited for high performance graphics and it has a PCI bus interface. The knowledge about the bus interface poses a constraint on the mainboard that can be used and therefore restricts the set of correct con gurations. Unlike transformational adaptation which transfers a previous solution to a new problem, generative adaptation transfers the derivation of a previous solution (the reasoning trace) to the new problem. Therefore, this kind of adaptation is also called derivational analogy [Carbonell, 1986]. Reasoning traces record information about the derivational process that lead to a particular solution in the past, including decision information, justi cations, and reasoning alternatives. During generative adaptation, these traces are essentially replayed in the context of the new target problem, i.e., the same or similar decisions are also tried to solve the current problem. If certain decisions from a previous reasoning trace cannot be transferred to the current problem, the generative problem solver will come up with a decisions by itself, without reusing information contained in the cases. In order to enable replay of reasoning traces, they must of course be stored in the cases, i.e., the representation of cases must be extended. We can basically distinguish two dierent kinds of replay strategies: one shot replay and interleaved replay (see Figure 2) which we will discuss in the following sections. One Shot Replay NEW PROBLEM

NEW PROBLEM

CASE

Problem ... Games: 10 Music: 0 ....

DR

Problem ... Games: 0 Music: 10 ....

Interleaved Replay

New Solution Trace 1. Select Motherboard (> ASUS) 2. Select CPU (>Pentium 200 MHz) 3. Select CD-ROM (>Sony 10x) 4. Select Graphics Adaptor (cheap S3) 5. Select Soundcard (>MIDI 634)

Fig. 3.

Solution Trace 1. Select Motherboard (> ASUS) 2. Select CPU (>Pentium 200 MHz) 3. Select Graphics Adapter (>Matrox) 4. Select Joystick (>JK 485) 5. Select CD-ROM (>Sony 10x)

Problem ... Games: 0 Music: 10 ....

New Solution Trace 1. Select Motherboard (> ASUS) 2. Select CPU (>Pentium 200 MHz) 3. Select Graphics Adaptor (cheap S3) 4. Select CD-ROM (>Sony 10x) 5. Select Soundcard (>MIDI 634)

Example for one shot and interleaved replay.

One Shot Replay In one shot replay, one rst identi es which portions of the solution trace can

be reused in the context of the new situation. These decisions are then replayed by the problem solver. After the replay is nished, the problem solver will make the remaining decisions on its own. The left side of Figure 3 shows an example in the PC domain. The new problem is to con gure a PC

T

basically used for music applications and not used for games. The retrieved case in the case-base, however, is con gured for being used for games and not for music. The depicted solution trace for the case shows the con guration steps that had been take to con gure this particular PC. In one shot replay, the decisions 1,2, and 5 from the case are replayed rst because they are also valid decisions for the new problems. Then, the remaining decisions concerning a dierent graphic adapter and the soundcard are taken. Caplan/Cbc [Mu~noz and Weberskirch, 1996] is an example of a case-based planning system that uses one-shot replay.

Interleaved Replay In interleaved replay, we can switch several times between the replay of

DR

AF

a previous solution trace and the generation of certain new decisions by the problem solver. The solution trace is followed up to a certain point where following it further would not be bene cial any more. Then, the generative problem solver takes over until a point is reached where the previous solution trace can be followed again. The right side of Figure 3 shows interleaved replay for the same problem used already for demonstrating one shot replay. First, decisions 1 and 2 are replayed. Then, the problem solver takes over and determines a dierent graphic adapter. Then, again, step 5 from the case is replayed, i.e., the CD-ROM is selected. Finally, the problem solver con gures the soundcard being used. Prodigy/Analogy [Veloso, 1994] is an example of a cased-based planning system that uses interleaved replay. We see that the basic dierence between the two approaches is only the sequence in which replay occurs and not the steps that are replayed. In this examples it might seem easy to decide whether a step can be replayed for a new situation or not. However, in many situations this is not the case, particularly when there is a lot of interaction between the reasoning steps. In interleaved replay, it can also become very dicult to decide when to switch between replay and problem solving. Up to now, there is no general solution to this problem.

Solution trace replay vs. complete decision replay An other important dierence between

dierent techniques is the amount of information about the solution trace that is reused. In solution trace replay, only the selection of the solution elements (selection of PC components in Figure 3) is recorded and replayed. In complete decision replay all reasoning information in the problem solving process are recorded and possibly take into account. These reasoning information can be justi cations, and reasoning alternatives, or failed attempts.

3.6 Compositional Adaptation

In addition to these adaptation models, recent research has demonstrated the power of delivering solutions through the retrieval, adaptation, and subsequent composition of multiple cases. This

AF

3.7 Hierarchical Adaptation

T

leads to compositional adaptation [Redmond, 1990, Sycara and Navinchandra, 1991], in which newly adapted solution components from multiple cases are combined to produce a new composite solution. Newer approaches indicate that it makes sense to compose a solution from parts of several old cases. This is possible if the solution consists of dierent parts which can be adapted more or less independently and it is eective if there are few con icts between these components so that a change in one component does not have several side-eects on other components. Compositional Adaptation is used in Cadsyn [Maher and Zang, 1993], Composer [Purvis and Pu, 1995], DeJ a V u [Smyth and Cunningham, 1992], Prodigy/Anology [Veloso, 1994] or Caplan/Cbc [Mu~noz and Weberskirch, 1996].

Hierarchical adaptation is an other recent development that is used in combination with the shown adaptation models [Bergmann and Wilke, 1996]. Cases are stored at several levels of abstraction and the adaptation is performed in a top-down fashion. At rst, the solution is adapted at the highest level of abstraction (omitting less relevant details). Then, the solution is re ned in a stepwise manner and the required details are added. Hierarchical adaptation can either reuse a single case or it can reuse dierent cases for dierent levels of abstraction or for re ning dierent details of the solution. Examples of systems using hierarchical adaptation are DeJ a V u [Smyth and Cunningham, 1992], Paris [Bergmann and Wilke, 1995], or the strati ed CBR approach by [Branting and Aha, 1995].

4 Discussion

DR

Adaptation is still a big search issue in case-based reasoning that is also getting increasingly important in practice as the complexity of applications increases. During the last years, there have been several workshops and published papers [Kolodner, 1993, Hanney et al., 1995, Vo, 1996] that try to systematically analyse the dierent approaches to adaptation that can be found in current CBR systems. However, all these approaches lack a systematic approach to design an appropriate adaptation method particularly suited for a given application and considering the knowledge available in this domain. We consider this one of the major challenges that need to be addressed in the future.

References

[Bain, 1986] Bain, W. M. (1986). A Case-Based Reasoning System for Subjective Assessment. In Proceedings of the 5th Annual National Conference on Arti cial Intelligence AAAI-86, pages 523{527, Philadelphia, Pennsylvania, USA. AAAI86, Morgan Kaufmann Publishers.

DR

AF

T

[Barletta and Hennessy, 1989] Barletta, R. and Hennessy, D. (1989). Case adaptation in autoclave layout design. In [Hammond, 1989], pages 203 { 207. [Bergmann and Wilke, 1995] Bergmann, R. and Wilke, W. (1995). Building and re ning abstract planning cases by change of representation language. Journal of Arti cial Intelligence Research, 3:53{118. [Bergmann and Wilke, 1996] Bergmann, R. and Wilke, W. (1996). On the Role of Abstraction in CaseBased Reasoning. In Proccedings of the EWCBR-96. [Branting and Aha, 1995] Branting, K. L. and Aha, D. W. (1995). Strati ed case based reasoning: Reusing hierarchical; problem solving eposides. Technical Report AIC-95-001, Naval Research Lab. [Carbonell, 1983] Carbonell, J. G. (1983). Derivational Analogy and Its Role in Problem Solving. In Proceedings of the 3rd Annual National Conference on Arti cial Intelligence AAAI-83, Washington, D.C., USA, August 1983. AAAI, Morgan Kaufmann Publishers. [Carbonell, 1986] Carbonell, J. G. (1986). Derivational analogy: a theory of reconstructive problem solving and expertise acquisition. In Michalski, R., Carbonnel, J., and Mitchell, T., editors, Machine Learning: an Arti cial Intelligence Approach, volume 2, pages 371{392. Morgan Kaufmann, Los Altos, CA. [Goel and Chandrasekaran, 1989] Goel, A. K. and Chandrasekaran, B. (1989). Use of Device Models in Adaption of Design Cases. In [Hammond, 1989], pages 100{109. [Hammond, 1989] Hammond, K. J., editor (1989). Proceedings: Case-Based Reasoning Workshop. Morgan Kaufmann Publishers. [Hanney et al., 1995] Hanney, K., Keane, M. T., Smyth, B., and Cunningham, P. (1995). Systems, tasks and adaptation knowledge: revealing some revealing dependencies. In Veloso, M. and Aamodt, A., editors, Case Based Reasoning Research and Development: First International Conference, ICCBR-95, pages 461{ 470. Springer, Berlin. [Hennessy and Hinkle, 1991] Hennessy, D. and Hinkle, D. (1991). Initial results from clavier: A case-based autoclave loading assistent. In Case-Based Reasoning, Proceedings of a Workshop, pages 225 { 232, Washington, D.C. DARPA, Morgan Kaufmann, San Mateo. [Kolodner, 1993] Kolodner, J. L. (1993). Case-Based Reasoning. Morgan Kaufmann, San Mateo. [Leake, 1996] Leake, D. B., editor (1996). Case-Based Reasoning: Experiences, Lessons, and Future Directions. AAAI Press, Menlo Park, CA. [Maher and Zang, 1993] Maher, M. L. and Zang, D. M. (1993). Cadsyn: A case-based design process model. In Process Model. Arti cial Intelligence for Engineering Design, Analysis, and Manufacturing, volume 7(2), pages 97{110. [Mu~noz and Weberskirch, 1996] Mu~noz, H. and Weberskirch, F. (1996). Planning for manufacturing workpieces by storing, indexing and replaying planning decisions. In Proc. 3rd Int. Conference on AI Planning Systems (AIPS-96). AAAI-Press. [Purvis and Pu, 1995] Purvis, L. and Pu, P. (1995). Adaptation using constraint satisfaction techniques. In Veloso, M. and Aamodt, A., editors, Case Based Reasoning Research and Development: First International Conference, ICCBR-95, pages 289{300. Springer, Berlin. [Redmond, 1990] Redmond, M. (1990). Distributed Cases for Case-Based Reasoning; Facilitating Use of

DR

AF

T

Multiple Cases. In AAAI, editor, Proceedings of the Annual National Conference on Arti cial Intelligence AAAI-90, pages 304{309, Boston, Massachusetts, USA. AAAI, Morgan Kaufmann Publishers. [Richter, 1995] Richter, M. M. (1995). The knowledge contained in similarity measures. Invited Talk on the ICCBR-95. http://wwwagr.informatik.uni-kl.de/~lsa/CBR/Richtericcbr95remarks.html. [Smyth and Cunningham, 1992] Smyth, B. and Cunningham, P. (1992). Deja Vu: a hierarchical case-based reasoning system for software design. In Neumann, B., editor, ECAI 92: 10th European Conference on Arti cial Intelligence, August 1992, Vienna, pages 587{589. Wiley, Chichester. [Smyth and Keane, 1996] Smyth, B. and Keane, M. T. (1996). Using adaptation knowledge to retrieve and adapt design cases. Journal of Knowledge Based Systems. [Sycara and Navinchandra, 1991] Sycara, K. P. and Navinchandra, D. (1991). In uences: A thematic abstraction for the creative reuse of multiple cases. In Proceedings of the Case-Based Reasoning Workshop, pages 133{144, Washington D.C., USA. [Thompson, 1997] Thompson, V. (1997). Corporate memories. Byte, International Edition, pages 32IS 7{12. [Veloso, 1994] Veloso, M. M. (1994). Planning and Learning by Analogical Reasoning. Number 886 in Lecture Notes in Computer Science. Springer, Berlin. [Vo, 1996] Vo, A. (1996). How to solve complex problems with cases. Engineering Applications of Arti cial Intelligence, 9:377{384. [Wilke and Bergmann, 1996] Wilke, W. and Bergmann, R. (1996). Considering Decision Costs During Learning of Feature Weights. In Proccedings of the EWCBR-96.

This article was processed using the LATEX macro package with LLNCS style