JUNE Artificial Intelligence Distance Learning Philosophy

JANUARY - JUNE 2003 25

IntelligenceLearning Philosophy

Support of PfP MandateKrunoslav An tolis *

Computers have long been utilised in the legal environment. The main use of computershowever, has merely been to automate office tasks. More exciting is the prospect of using

artificial intelligence (AI) technology to create computers that can emulate the substantivelegal jobs performed by lawyers, to create computers that can autonomously reason with the

law to determine legal solutions, for example: structuring and support of Partnership forPeace (PjP) mandate. Such attempts have not been successful jet. Modelling the law and

emulating the processes of legal reasoning have proved to be more complex and subtle thanoriginally envisaged. The adoption by AI researchers specialising in law of new AI techniques,such as case based reasoning, neural networks, fuzzy logic, deontic logics and non-monotoniclogics, may move closer to achieving an automation of legal reasoning. Unfortunately these

approaches also suffer several drawbacks that will need to be overcome if this is to beachieved. Even if these new techniques do not achieve an automation of legal reasoning

however, they will still be valuable in better automating office tasks and in providing insightsabout the nature of law.An idea to apply the technology of intelligent multi-agent systems tothe computer aided learning (CAL) in law, is currently being developed as a research projectby the author of this article (see e.g. [Antolis, 2002.]). Similar projects are usually based onthe most modern technologies of multimedia and hypermedia, as it was implemented in thisarticle. The theoretical foundations of the design and architecture of intelligent system for

decision support process in law and distance-learning environment are, however,at their early stage of development.

Key words: intelligent systems, law, legal expert systems, decision support process,distance-learning environment, legal reasoning, mandate, multimedia, hypermedia

ArtificialDistance

•l.n

1. Introduction

Computers have long been utilised in thesphere oflaw. Basic applications such as word proc-essors, spreadsheets and databases have all foundtheir way into legal offices. Recently, more sophisti-cated tools such as computerised legal research sys-tems, document drafting packages, and practice man-agement systems have become increasingly common.Most exciting however, has been the prospect of us-ing artificial intelligence (AI) techniques to create

. Krunoslav Antolls, PhD, Assistant professor, Department forBusiness Computing, Faculty of Economy, University of Zagreb,Croatia

&

Original paperUDC 007:34:327

Received in Octobre 2002

'automated legal reasoning systems', computer sys-tems that reason with and apply the law in an effortto resolve legal disputes. Examples of such systemsinclude legal expert systems.

However, the practical benefits of such auto-mated reasoning systems have fallen short of earlyoptimistic predictions; they have not resulted in com-puter systems that can independently and inexpen-sively provide expert advice about substantive law.

The failure of efforts to create automated le-gal reasoners has led to a reassessment and reclassi-fication of research aims. The key objective has beentransformed into the creation of 'knowledge basedsystems'. The purpose of a knowledge based system

26 CROATIAN INTERNATIONAL RELATIONS REVIEW

is not necessarily autonomously to provide solutionsto legal problems. Instead, the goal is to incorporatelegal knowledge and reasoning strategies into theautomation of legal tasks so as to make the systemsthat perform those tasks more productive. I This goalis far less grand than that for automated legalreasoners.

In addition to this change of focus, new ap-proaches in AI have recently emerged. These newapproaches may overcome some of the limitationsinherent in earlier attempts to automate legal rea-soning. What general benefits these new approachesmight provide, and what specific benefits knowledgebased systems might provide in law is open to ex-amination.

The interest in the systems of intelligent sys-tem for decision support process in law and distancelearning environment (DLE) in law, and self-learn-ing in general, is stimulated by the growing popular-ity and easy access to the World Wide Web. The ex-pressive capabilities of HTML and its support bythe popular browsers makes it available to provideon-line courses of deep complexity. It is generallyagreed (see e.g. [Gorodetsky, 1996; O'Hare andJennings, 1996]), that at least three important issuesshould constitute the basis for such learning systems:they are expected to be distributed, intelligent andadaptive. The adaptability of IDLE systems, in allmeanings of this concept, is announced in most ofcurrently developed projects, but usually weaklyimplemented in practice. Adaptation and learning inmulti-agent systems establishes a relatively new butsignificant topic in Artificial Intelligence (AI). Multi-agent systems typically are very complex and hardto specify in their behaviour. It is therefore broadlyagreed in both the Distributed Al and the relatedcommunities that there is the need to endow thesesystems with the ability to adapt and learn, that is, toself-improve their future performance. There maybe at least two different kinds of adaptation: purelyengineering one, related to the situation, as mentionedabove, when a program is able to change its behav-iour in changing context regardless of the nature ofthe latter; and one arising from Human-ComputerInteraction (HCI) issues (see e.g. [Gavrilova andVoinov, 1995]), which concern a program adjustingits behaviour to meet the user's expectations andpeculiarities according to some reasonable strategy.

The second, HCI-related kind of adaptation iscrucial for any tutoring system. It is evident, that aqualified tutor works differently with different stu-dents, regardless of possibly equal "marks" of thelatter. This issue is even more important for distancelearning systems, in which the online contact withhuman tutor is lost.

Therefore, one of the main directions of thedescribed project is to investigate possible strategiesof the automatic adaptation of a DL system, basedon the concept of Student Model (see e.g. [Gavrilovaand Voinov, 1995; Gavrilova et al., 1996]) and thecorresponding Learning Process Model. In contrastto cited papers, these models are supposed not tofunction separately, but to form an "agent model" inthe meaning of (see e.g. [Shoham, 1993]). These in-vestigations are accompanied by the building of adap-tive DL system IDLE (Intelligent Distance LearningEnvironment) that will be implemented for the de-velopment of adaptive courseware on WWW in thefields of intelligent system for decision support proc-ess in law and law sciences.

Before examining these questions however, itis worthwhile exploring the failure of early attemptsto automate legal reasoning.

2. Automated legal reasoning

As we have seen, early attempts to automatelegal reasoning involved the creation of legal expertsystems. Expert systems contain knowledge and rea-soning strategies such that the computer can applythe knowledge to problems in order to determine asolution. However, these early legal expert systemstended to be developed with scant regard for juris-prudence (see e.g. [Susskind 1987]).

In an expert system, knowledge is stored inthe system in the form of 'production rules'. Theserules take the form

If (condition) then (result)For exampleIf (mens rea for murder) and (actus reus for

murder) then (guilty of murder)reasoning in such production rule systems is a

process of applying the rules stored in the system,by satisfying the clauses of the production rules. Thisrepresentation of legal knowledge has a superficialappeal if one believes that the law is no more than aseries of defined rules which lawyers and judges sim-ply apply to a problem to determine a solution. While,superficially, this may appear to be what occurs whenlawyers reason about the law, this view of law is, infact, only subscribed to by extreme legal positivists(see e.g. [Zeleznikow & Hunter 1994]).

The jurisprudentially assumptions inherent inthe 'production rule' approach to modelling legalknowledge results in numerous difficulties. How doesone determine whether a situation is within the am-bit of a rule? What happens if no clear rule can befound that governs the situation? What if rules con-flict? In common law legal systems these and other


similar problems are, at least partially, resolved byreference to previously decided cases; a process for-malised in the doctrine of stare decisis.

Prima facie, reference to and reasoning fromcases presents problems for production rule systems.Such reference requires the legal expert system toincorporate, at the least, analogical reasoning." Pro-duction rule systems can only incorporate an ap-proach based on deductive reasoning. As such, thisseverely limits the ability of production rule systemsto model the law and emulate legal reasoning. Earlylegal expert systems failed to achieve their goal be-cause they did not account for the complexity andsubtlety of the law and of legal reasoning.

3. Alternatives to production rules

In the quest to create intelligent computer sys-tems, AI researchers have investigated numerous al-ternatives to production rule expert systems. How-ever, those researchers specialising in AI and lawapplications have, so far, only explored a few of thesetechniques. Those that have been examined include:case based reasoners; artificial neural networks (neu-ral nets); fuzzy logics, deontic logics and non-monotonic logics (see e.g. [Zeleznikow & Hunter1994)). To determine whether any of these ap-proaches can aid in the creation of automated legalreasoners, we must first examine the jurisprudentiallyassumptions inherent in the way that law is mod-elled with each of these alternatives.

4. Case based reasoners

In stark contrast to the methods used to modellegal knowledge adopted in production rule systems,

Figure 1.

Stored case

case based reasoners do not rely on there being rulesof law. Instead, in a case based reasoning system, asthe name implies, one or more individual examplesare stored by the system.' The case based reasonerthen manipulates the cases in its knowledge base soas to reach a conclusion when given a concrete prob-lem. In this way case based reasoners have been cre-ated that not only argue to a solution, but also presentan 'argument' as to why the solution should occur,highlight weak points in the argument presented andindicate what likely responses the opposition willraise.

Capturing legal knowledge in this way hasintuitive appeal considering the importance of casesand judge made law contained in cases, in commonlaw legal systems. In applying cases and construct-ing arguments from cases, case based reasoners ap-pear to incorporate analogical reasoning. Thus, po-tentially, they appear able to overcome some of theproblems that plague production rule based attemptsto automate legal reasoning.

Unfortunately this has not proved completelytrue. To be able to reason analogically, a person orcomputer system must be able to decide when twocases are similar enough for an analogy to be cre-ated between them (see e.g. [Mital and Johnson1992]). How is this done? The mechanics of thisprocess are uncertain in humans. Researchers infields such as psychology and cognitive science givediffering accounts. Whilst theories abound, none isoverwhelmingly accepted (see e.g. [Mital andJohnson 1992]).

In case based reasoners, the process of simi-larity determination and analogy formation occursby breaking down a situation into 'factors'. Thesefactors are pre-determined by the programmer of thecase based reasoning system. When presented witha problem situation, the case based reasoning sys-

Problem situation

t-actor I t-actor IFactor 2Factor 3 Factor 3f=;lrt()r 4 f=;lrt()r 4

Factor 5

Factor N Factor N


tern also breaks it down in terms of these factors (seee.g. [Ashley 1989]). In this way, the case and theproblem situation can be compared for the presenceor absence of factors. The number of factors sharedresults in a measure of how similar one case or situ-ation is to another,"

From here analogies can be created and ma-nipulated. Determining in what ways the cases aresimilar and in what ways they differ in turn allowsthe construction of likely arguments for and againsta given proposition.

In this way, case based reasoners appear toemulate legal analogical reasoning; cases are com-pared with each other and the similarities and differ-ences between them are used as the basis for the con-struction of arguments. However, as will be arguedsubsequently, what such systems actually implementis only a crude approximation of analogical reason-mg.

5. Neural nets

Neural nets operate in a similar, but subtlydifferent way to case based reasoners.' Like casebased reasoners, neural nets do not operate with le-gal rules, but with legal cases.

A neural net is a computer model that is in-spired by and that mimics the structure of a biologi-cal nervous system, particularly the human brain. AIresearchers hope that by mimicking the underlyingstructure of an undeniably intelligent system, theywill also be able to emulate the intelligent behaviour

Figure 2.

Factor 1

of that system. In computer science, neural nets haveproved exceptionally good at many tasks which haveconfounded expert systems; such as computer vision,speech recognition and handwriting recognition.Notably, all these tasks involve the recognition andclassification of patterns, a by-product of the factthat neural nets are inherently suited to tasks involv-ing recognising and classifying patterns.

Identifying patterns between circumstancesand classifying those patterns ostensibly appears tobe what is occurring when an analogy is formed. Twocases will only be regarded as similar if they sharesome sufficiently close pattern of facts. Researchersin AI and law have thus viewed neural nets as a prom-ising technique with which to emulate legal analogi-cal reasoning; their ability to match patterns couldbe used to match cases with each other and with prob-lem situations.

Unfortunately, like cased based reasoners theuse of neural nets is problematic. This will be dis-cussed shortly.

6. Fuzzy logic based systems

Another alternative to the production rule sys-tems which we discussed above are fuzzy produc-tion rule systems; systems that use fuzzy logic. Incontrast to the classical logic on which traditionalexpert systems are based, fuzzy logic is said to beable to deal with imprecision and partial truths. Theoriginator of fuzzy logic, Zadeh, thought fuzzy logicwould be able to model the imprecision inherent in

Factor 2 Factor 3 . Factor N

Input

Output

Classification


'linguistic variables'. Law is replete with such lin-guistic variables, making it a seemingly attractivefield for the application of fuzzy logic. However, theuse offuzzy logic in law has been criticised by someAI and law researchers who claim that its philosophi-cal basis is uncertain when applied to legal concepts(see e.g. [Bench-Capon & Sergot 1988]). It does seemdoubtful whether merely using fuzzy logic to extendproduction rule systems provides much benefit. Thisdoes not address many of the underlyingjurisprudentially problems which are inherent in pro-duction rule systems; notably the reliance on deduc-tive reasoning. However, the potential benefit of in-corporating fuzzy logic into case based reasoners andhybrid reasoning systems (systems that incorporatemultiple reasoning strategies) remains to be exploredin detail.

7. Deontic logic based systems

Unlike traditional logic, fuzzy logic, casebased reasoners or neural nets, deontic logic is con-cerned with explicitly modelling the normative as-pects oflaws. In traditional production rule systems,laws are merely 'if ... then ... ' statements with no ex-plicit moral content. However, it is argued that it isnecessary to account for the normative aspect of lawif the law is to be successfully modelled (see e.g.[McCarty 1989]). Deontic logic aims to explore therelationships between normative aspects of the law.

Attempts to utilise deontic logic to create au-tomated legal reasoning systems are at an early stageand have not been completely successful. The use ofdeontic logic has also endured weighty criticism.However, like approaches based on fuzzy logic, theuses of deontic logic remain to be fully explored andthe future benefit that such approaches may provideremains as yet unpredictable.

8. Non-monotonic logic basedsystems

A feature of traditional logic, fuzzy logics, anddeontic logics is that once a statement is found to betrue under the system, the addition of new facts orstatements to the system will not alter that earlierfinding. This is known as monotonicity. However,an important feature of human reasoning is its con-tingent nature; results are always subject to falsifi-cation by new information. Emulating this aspect ofhuman reasoning is desirable. Non-monotonic rea-soning systems attempt to do this. In contrast tomonotonic reasoning systems, the addition of new

facts or statements to a non-monotonic reasoningsystem can alter earlier findings. The contingent na-ture of non-monotonic logics thus seems a promis-ing addition to techniques for describing legal knowl-edge. Like fuzzy logic and deontic logics however,the use of non-monotonic logics in automating legalreasoning has not been comprehensively exploredand so its benefits are uncertain.

Thus, while case based reasoners, neural netsand the various logics discussed above provide al-ternatives to production rules, AI and law research-ers have only explored the use of case based reasonersand neural nets in any detail. The application of casebased reasoning and neural net techniques is, how-ever, subject to several problems.

9. Problems with case basedreasoners and neural nets

At first glance, both case based reasoners andneural nets appear extremely attractive for use inautomated legal reasoning systems; their primafacieability to emulate analogical reasoning should allowthem to overcome some of the problems experiencedwith production rule systems. Unfortunately, this hasnot proved unequivocally true. While both case basedreasoners and neural nets can match similar caseswith each other, this is achieved at a low level ofcomplexity. As previously mentioned, similarity isonly found by testing for the presence or absence ofpredefined factors.

While researchers may not know all that isinvolved in the process of analogising by humans,even a superficial examination shows that the proc-ess involves more than the mere comparison of thepresence or absence of predefined factors. Any simi-larity perceived between situations depends on thecontext in which the situations are viewed. The veryfactors that are considered relevant in the finding ofsimilarity is dependent on the context under whichthe comparison is being made (see e.g. [Tito 1987]).Thus the ability to create an analogy, is determinedby the context under which the similarity is deter-mined.

Just as importantly, the process of analogis-ing depends not only on linking situations as simi-lar, but also moulding the analogy made in a direc-tion desired for the outcome of an argument. Againhowever, the way that an analogy is manipulateddepends on its context, on one's viewpoint and theoutcome that one wants to achieve. Thus, as severaljurisprudes note, analogising is inherently depend-ent on the overall purposes that the legal system istrying to achieve (see e.g. [MacCormick 1978]). The


classification of facts for the purposes of creatinganalogies occurs in a whole body of knowledge andtheory we use to make sense of the world. Whendeciding between competing fact classifications andcompeting arguments, our evaluation inherently in-volves considering the consequences of each out-come on our model of the world. In this sense, simi-larities, dissimilarities, classifications and thus theexistence of analogies is made and not found.

For these reasons, breaking cases and prob-lem situations into predefined categories is neces-sarily a far less subtle method of analogy makingthan that which occurs during human legal reason-ing. Even disregarding any of the practical and tech-nical difficulties in doing SO,6 while case basedreasoners and neural nets can mimic analogical rea-soning, the result is crude.

Figure 3. The architecture of the DL system IDLE

.:!r

V

!

..Dc:anacilt Kn:)w~e

10. The Architecture of IntelligentDistance Learning Environment

The software implementation of the IDLE pro-totype may be regarded as: multi-agent, portable,access-restricted and using multimedia effects. Sys-tem contains both the modules of traditional (iso-lated) architecture, which examples are tutor's andsystem administrator's workbenches, and the mod-ules of multi-agent architecture, which implementse.g. system supervising, immediate control overlearning. Among the agents comprising the system,are those, which may be regarded as "classic" (seee.g. [Takaoka, Okamoto, 1997]):

- Expert Tutor. Expert Tutor controls thelearning process, applies different education strate-gies according to the cognitive and essential personalfeatures of the student, together with his/her educa-tional "progress". Interacts with Interface Engineeron adaptation and user modelling issues.

I~

Krow~eonTu~

"

'< ..••...........•... .........

IDLE:Intelligent Distant Lflarning

Environmeni

( WVIVv-hrowser •

Cif3(~ .~StuJelll-A S\u<kra-B

t

StlldA1'lt -1;


- Interface Engineer. Interface Engineermaintains User Model for current student (both bythe preliminary testing and dynamically during thestudent's working with the system) and provides itto the human users and to the other agents. In par-ticular this information is used in choosing the sce-nario of the presentation of courseware and evaluat-ing the student progress and his/her assessment.

- Domain Expert (Domain Assistant). Do-main Expert accumulates and provides knowledgeon subject domain, maintains testing exercises, pro-vides these data to its users and other agents, analy-ses the students "feedback" in domain terms.

Such system architecture may be regarded asa development of a traditional for AI idea of "activeknowledge bases": each of the knowledge base com-ponents is access via corresponding agent, whichadapts knowledge to the context of usage. Further-more, a close analogue of the agent-mediated knowl-edge bases is seen in the works of [van Harmelin,1992] where a knowledge base with reflection wasdeveloped. It has both domain knowledge and somemeta-knowledge about its integrity, and could applythis in reasoning. Certainly some close analogies ofthese techniques are seen also in the field ofdatabases, especially object-oriented ones [Cattell,1991] .

11. Adaptive tutoring

IDLE's adaptability is twofold. The main isscenario adaptability which is controlled by experttutor who produced different hypertext navigationrules for different students according their user/stu-dent models. In contrast to the traditional mode ofhypertext navigation, when the default routes areembedded into the page, in the described system thenavigation is guided by the expert tutor, who restrictsthe navigation freedom, and puts "barriers". It doesnot allow student to move to those places of course,where his/her visit is unreasonable (according to thestrategy and the target of learning). For this, all thedomain material is stratified into "clusters", contain-ing one or more hypertext nodes and representing"learning units". The set of clusters, which nodeswere visited by the student during all his/her learn-ing sessions, is called "scope of visibility" and oc-curs as a union of both already and currently learnedmaterial. The navigation within the scope of vi sibil-ity is unrestricted. To widen the scope of visibility,the student should overcome a "barrier", which pro-vides some examination test. The interface adapt-ability is guided by the interface engineer who may

change the main interface parameters according tothe individual students' attitudes and preferences[Gavrilova, Voinov, 1995].

12. Conclusion

Given all the limitations discussed above, onemight conclude that AI has nothing to offer lawyers.However, it is the author's view that this would be ahasty conclusion. .

Whilst the early hopes of researchers in AIand law - the creation of automated legal reasoningsystems - have not have been fulfilled, such effortshave been far from fruitless. While case basedreasoners and neural nets are also incapable of pro-viding realistic automated legal reasoning systems,these approaches do provide a more sophisticatedemulation oflegal reasoning than do production rulesystems.

However, with the refocusing of effort intocreating legal knowledge based systems, the benefitsof applying AI techniques to the law become moreapparent. Once the goal is no longer primarily to cre-ate automated legal reasoners, but rather to build le-gal knowledge based systems, even the productionrule approach to computerising legal knowledge maystill be useful. The utilisation of case based reason-ing and neural net approaches provide additionalbenefits.' The incorporation of production rules, casebased reasoners, neural nets and perhaps alternativelogics into hybrid knowledge based systems holdspromise. The introduction of these AI techniques intoexisting computer applications may greatly improvethe sophistication and utility of such existing sys-tems. Either approach may lead us to exciting appli-cations yet to be envisaged.

So, while the goal of producing a truly intelli-gent computer system that can automatically reasonwith law is beyond the realms of possibility atpresent, research continues. Even if this holy grail isnever reached, the insights gained may, none-the-less, still provide immediate great benefits to law-yers and the legal system.

And at last but not least, we try to build andsuggest the perceivable architecture of intelligentsystem for decision support process in law and dis-tance learning environment, for developing such sys-tem in future. •

32

NOTES

BIBLIOGRAPHY

1 An expert system whose aim is merely to provide advice andguidance to the user, and not autonomously to provide reasonedsolutions is essentially a knowledge based system. To this ex-tent, such expert systems are a subset of knowledge based sys-tems. Thus the distinction drawn is not a concrete one, the dif-ference between the two merely focusing on the task that thesystem is trying to perform. For the purposes of this discussion,the term 'expert system' will be used to refer to systems (basedon production rules) designed to independently provide solu-tions to substantive legal problems.

2 Attempts have been made to develop production rule systemsto reason with cases e.g. through the use of rule induction sys-tems. However, it is argued that such approaches have prob-lems of their own, most importantly the loss of richness andplasticity in the representation of the case based knowledge.Further, given the view that analogical reasoning plays a highlyimportant role in legal reasoning, the emulation of this form ofreasoning by a legal expert system seems desirable; rather thanthe simple extension of deductive reasoning.

3 The term 'case based reasoner' was coined by researchers incomputer science and not law. Hence the term 'case' does notgenerally refer to the notion of a legal case in which there aretwo opposing parties in dispute. Instead the idea of a case ismore general, it is simply an example. Thus the techniques usedcould equally be denoted by the terms 'example based reason-ing' or 'experience based reasoning'. Just to confuse mattersthough, when case based reasoners are applied in law the 'ex-amples' they use are generally actual legal cases! It is howeverimportant to be aware of the general distinction. In this discus-sion, a case based reasoner will refer to an example basedreasoner in which the examples used are actually legal cases.

4 This is of course a simplification. Often systems compare notonly the number of factors that are shared, but also attempt totake account of their relative importance within the case or prob-lem. While such strategies provide additional subtlety, the crudecomparison of factors remains the backbone of current casebase reasoning approaches to analogical reasoning.

5 In a case based reasoner, the relative importance of factorsand the relations between those factors have to be determined

1. Antolis K. (2002) "Dr. Antolis' presentation aboutCroatia's approach to develop an ADL system", Workshop ofthe PfP Consortium, Advanced Distributed Learning - WorkingGroup (ADL-WG), 4-6 of September, 2002 in Kiev.(www.isn.ethz.ch/adl-wg/home.htm)

2. Ashley, K 0 (1989) "Toward a Computational Theoryof Arguing with Precedents: Accommodating Multiple Interpre-tations of Cases", Proceedings of the Second International Con-ference on Artificial Intelligence and Law (ACM, New York).

3. Bench-Capon, T J M & Sergot, M J (1988) "Toward aRule-Based Representation of Open Texture in Law" in Walter,C (ed) Computer Power and Legal Language (London: Quorom).

4. Cattell R.G.G (1991). Object Data Management.Addison-Wesley.

5. Gavrilova T., Voinov A., Zudilova E. (1995) User mod-elling technology in intelligent system design and interaction.In: Proceedings of East-West International Conference on Hu-man-Computer Interaction HCI'9S, Moscow,115-128.

6. Gorodetsky V.1. (1996) Multi-agent systems: currentstate of research and the application perspectives. In: Proc. VNat.AI.Cont. KII-96. pp. 36-45. (in Russian.)

7. McCarty, L T (1989) "A Language for Legal Discourse:I. Basic Features", Proceedings of the Second InternationalConference on Artificial Intelligence and Law (New York: ACM).

8. MacCormick, N (1978) Legal Reasoning and LegalTheory (Oxford: Oxford University Press).

CROATIAN INTERNATIONAL RELATIONS REVIEW

by the designer of the system. In contrast, a neural net 'learns'these during its 'training'. Further, case based reasoners can'explain' how a particular analogy was reached. Neural netsmerely present an opaque conclusion. While extremely impor-tant in practice, these differences are not significant for thepresent discussion.

6 Production rules systems are difficult to build and maintain,there is a 'knowledge acquisition bottleneck'. The system's crea-tors have to explicitly code every rule and predicate manipu-lated by the system and then 'debug' them to ensure that theyare free of errors and inconsistencies and operate as predicted.Any changes have to be incorporated through the same time-consuming process. To make these systems 'intelligent' requiresmuch work by a domain expert working in conjunction with aknowledge engineer. Similar practical difficulties face the crea-tion of case based reasoners. In contrast, a neural net is 'trained'.Cases are repeatedly presented to the network; which 'learns'the case. This is an automatic process. This apparent ability toovercome the knowledge acquisition bottleneck makes the useof neural nets even more attractive in intelligent knowledge basedsystems. However, the training of neural nets creates furtherproblems. While the ability to learn information is attractive, thisalso has undesirable consequences. Firstly, when a neural net-work is presented with inconsistent cases it does not store thetwo cases separately, but rather stores a compromise betweenthose two cases. This approach may not be jurisprudentially jus-tified and could result in problems in the creation and manipula-tion of analogies. Further, neural nets rely on a form of statisti-cal analysis on the material that they are trained with. Conse-quently, if there is a lack of data then the training may be spuri-ous. This is a serious impediment to the use of neural-nets.

7 Further, the possibility of creating hybrid case based reasonersthat themselves incorporate neural nets has not been discussed.It is the author's belief that neural nets can be of benefit in sucha hybrid case based reasoner. Neural nets would be used at a'lower level' than currently, to determine similarity between situ-ations, while the manipulation, explanation and justification as-sociated with analogical reasoning would occur through othersystems.

9. Mital, V & Johnson, L (1992) Advanced InformationSystems for Lawyers (London: Chapman & Hall).

10. O'Hare G., Jennings N. (1996) eds. Foundations ofDistributed Artificial Intelligence. Sixth-Generation ComputerTechnology Series. Branko Soucek, Series Editor. John Wiley &Sons.

11. Shoham Y. (1993) Agent-oriented Programming.Artificial Intelligence, vol. 60, pp. 51-92.

12. Staniford G., Paton R. (1995) Simulating AnimalSocieties with Adaptive Communicating Agents. In: WooldridgeM. Jennings N.R. (eds) Intelligent Agents - Theories,Architectures, and Languages. Lecture Notes in Artificial Intelli-gence Volume 890. Springer-Verlag.

13. Susskind, R (1987) "Expert Systems in Law: A Ju-risprudential Inquiry" (Oxford: Clarendon Press).

14. Tito, C (1987) "Artificial Intelligence: Can Comput-ers Understand Why Two Legal Cases are Similar?" 7 Compu-ter/Law Journal 409

15. Zeleznikow, J & Hunter, 0 (1994) Building IntelligentLegal Information Systems: representation and reasoning in law(Dordrecht: Kluwer)

16. Takaoka R., Okamoto T. (1997) An Intelligent Pro-gramming Supporting Environment based on Agent Model. IEICE

Trans. INF & SYST., vol. E80-D, No.2 February.

Documents

JUNE Artificial Intelligence Distance Learning Philosophy