630 US ISSN 0271·9916 March 1988 RESEARCH EXTENSION SERIES 089

Expert Systems in Agriculture: Determining LimeRecommendations for Soils of the

Humid Tropics

Russell Yost, Goro Uehara, Michael Wade, M. Sudjadi, I P.G. Widjaja-adhi, and Zhi-Cheng Li

HITAHR . COLLEGE OF TROPICAL AGRICULTURE AND HUMAN RESOURCES . UNIVERSITY OF HAWAII

The Library of Congress has catalogued this serial publicationas follows:

Research extension series / Hawaii Institute of Tropical Agri­culture and Human Resources.-001--[Hono1ulu, Hawaii]:

The Institute, (1980-'v. : ill. ; 22 em.

Irregular.Title from cover.Separately catalogued and classified in LC before and

including no. 044.ISSN 0271-9916 = Research extension series - Hawaii

Institute of Tropical Agriculture and Human Resources.1. Agriculture-Hawaii-Collected works. 2. Agricul­

ture-Research-Hawaii-Collected works. I. HawaiiInstitute of Tropical Agriculture and Human Resources.II. Title: Research extension series - Hawaii Institute ofTropical Agriculture and Human ResourcesS52.5.R47 630'.5-dc19 85-645281

AACR 2 MARC-SLibrary of Congress [8506]

THE AUTHORSRussell Yost is an associate researcher, Department of Agronomy and Soil Science, College of TropicalAgriculture and Human Resources, University of Hawaii.

Goro Uehara is a soil scientist,. Department of Agronomy and Soil Science, College of TropicalAgriculture and Human Resources, University of Hawaii.

Michael Wade is a soil scientist, formerly with the Department of Soil Science, North Carolina StateUniversity.

M. Sudjadl is the director, Centre for Soils Research, Bogor, Indonesia.

I P. G. Wldjaja-adhl is the coordinator, Tropsoils Research Program, and a soil scientist, Centre forSoils Research, Bogor, Indonesia.

Zhl-Cheng Ll is a knowledge engineer, Tropsoils Project. Department of Agronomy and Soil Science.College of Tropical Agriculture and Human Resources, University of Hawaii.

This work was done under the auspices of the Tropsoils Project, comprising the Centre for SoilsResearch, Bogor, Indonesia; the University of Hawaii; and North Carolina State University.

CONTENTSPage

Introduction 1Expert Systems 1How We Chose Our Example 2Our Experience with EXSYS 3Future Applications of Expert Systems 6Literature Cited 8

EXPERT SYSTEMS IN AGRICULTURE: DETERMINING LIMERECOMMENDATIONS FOR SOILS OF THE HUMID TROPICS

Russell Yost, Goro Uehara, Michael Wade, M. Sudjadl,I P. G. Widjaja-adhi, and Zhl-Cheng Li

INTRODUCTIONExpert systems have recently attracted the

attention of agricultural scientists for applica­tion in a variety of information developmentand transfer situations. These computer soft­ware systems are designed to simulate one ormore of the ways that a human expert uses his orher knowledge and experience in making adiagnosis or a recommendation. While originalapplications included the diagnosis of bacterialdiseases (Hayes-Roth et al., 1983), useful applica­tions in agriculture include soybean diseasediagnosis (Michalski et al., 1980, 1982), manage­ment systems for apples, and the taxonomicidentification of turfgrass.

We have developed a prototype expert systemto make lime recommendations in the humidtropics. The objectives were to:

1. Document current methods of determininglime requirements for highly weathered soils ofthe tropics. This objective was developed as partof the Tropsoils/Indonesia project, which isadapting and developing lime recommendationtechnology for the highly weathered aciduplands of Sumatra, Indonesia.

2. Provide a way of transferring currentTropsoils research within Indonesia for use byextension workers and others with limitedagronomic training.

3. Provide an exploratory learning exercisefor ourselves about how an expert system is builtand what the potential applications might be.

EXPERT SYSTEMSTo exploit the potential of expert systems, we

must review what experts do. This includes(Michaelson et aI., 1985):

1. Applying their expertise to solve problemsin an efficient manner.

2. Explaining and justifying what they do.3. Communicating with other experts and

acquiring new knowledge.4. Restructuring and reorganizing knowl­

edge. Some individuals cannot accommodatemassive changes in their knowledge. However,experts should be able to restructure informa­tion based on new data and concepts.

5. Breaking rules. In certain fields of sciencethere are almost as many exceptions as there arerules. Experts understand both the spirit and theletter of the rule and are not bound by strict,literal interpretations of the concepts.

6. Determining relevance. They know when a

problem is clearly outside of their expertise andthat it should be referred to another expert.

7. At the boundary of their expertise,indicating whether their .information is notlikely to be the best available, and providingtheir information together with probablesources of better information.

Current expert systems partially achieveonly functions 1 and 2. It is quite likely, how­ever, that several of the other functions soonwill appear in newer expert system developmenttools.

Computer-based fertilizer recommendationshave been used successfully for many years.Users of such programs had to accept programrecommendations on faith or persevere inreading arcane FORTRAN code. Checking theprogram's rationale for soil samples that weregiven unreasonable recommendations was verydifficult, if not impossible, for those who didn'tread FORTRAN. An expert system-basedfertilizer recommendation would offer analternative to this difficulty. The user would beable to examine the rationale at any part in theprogram. This is because an expert system isdesigned not only to provide a recommendation,but to explain. how the recommendation wasdeveloped and, depending on the implementa­tion, to support the recommendation withliterature citations. This means there is usuallyan opportunity for the nonexpert to learn therationale. The emphasis of expert systems is toapply knowledge and information and to sharethat knowledge with others.

To adequately develop expert systems, onemust also learn techniques of knowledgeacquisition (Hayes-Roth et aI., 1983). Thisrequires extracting information from theexperts and representing it in a data base.

Extracting Information from the ExpertsIt is Widely recognized that knowledge

extraction is one of the more difficult problemsin developing expert systems. One technique isto have someone pose the questions and thesituations as realistically as possible so that theexperts can respond. It is well known thatexperts are better able to apply knowledge andexpertise than to explain and teach it.Knowledge extraction efforts might benefit fromsome of the experience and techniques describedby Kadane et al. (1980) in eliciting the priorinformation for Bayesian decision analysis.

1

Representing Knowledge In a Data Base orKnowledge Base

Here there are many approaches. such asrule-based systems. logic-based systems. andframe-based systems (ACM. 1985).

Rule-based systems. The most popular. rule­based systems use a sequence or chain of rules.These rules generally have the form:

IF1.2.

THEN1.2.3.orConclusion 1Conclusion 2

It is possible for the conclusion of one rule toserve as an IF condition for a subsequent rule. Inthis way rules are linked in a logical sequencethat simulates reasoning.

There appear to be three types of rules withwhich the system "reasons" or performs alogical sequence of questions and answers(Clancey. 1983):

1. Strategy rules. used to represent the planfor ordering the questioning or the presentationof the hypothesis and goals.

2. Structural rules. keys used to index theknowledge or as a way to reference a particularset of rules for a particular set of conclusions.

3. Support rules. those that supplement oradd conditions to the main conclusions. Theserules are easily changed or replaced because theydo not have major structural implications.

Logic-based systems. The second type ofknowledge representation is most fully devel­oped in the programming language PROLOG.PROLOG represents a fifth generation approachto problem representation and solution. As istypical of fifth generation "declarative"languages. the user first states the problem. Theproblem-solving algorithms of predicate calcu­lus then perform sophisticated matches andsubstitutions. The result is a vezy powerful logic­based derivation of conclusions. Languages suchas PROLOG can prove mathematical theoremsthrough their implementation of logic. It is alsopossible to develop rule-based expert systems inPROLOG that exploit the logic-based reasoningability. Such languages also permit machinelearning whereby a system can add rules to itselfdUring execution. To do this in a useful manner.however. the system must have meta-rules withinstructions on when and how to formulate thenew rules.

2

Frame-based systems. The third type ofknowledge representation has been available asexpert systems development software costing$50.000 and upwards. Frame-based representa­tion is explicitly developed to representhierarchical knowledge. In this case informa­tion is passed or transferred to an object becauseit belongs to a larger class of objects with astandard set of attributes. A frame is thestructure whereby attributes of an object arerecorded. including its belonging to anotherclass of objects. An example might be a soybeanplant. The frame would be the name "soybean"and all the attributes and characteristics of thesoybean. including the fact that it belongs to theclass of legumes. flowering plants. and plantsrather than animals. Becau~e the soybean is aplant. we know that it requires nutrients. water.light. and other inputs for growth. Because it is alegume. we know that there are rhizobia require­ments and other conditions for the rhizobia tobe effective if soil N is low. Many things can beinferred by simply knowing that the object is aplant and that it is a soybean plant. Thisinformation can be used by a frame-basedsystem to ask relevant questions or to developpatterns of reasoning. As suggested in the logic­based systems. hybrid systems can be developedthat use frames as units of information fromwhich rules can be formed.

HOW WE CHOSE OUR EXAMPLEAmeliorating the effect of soil aeidity was

selected as an example from agronomy and soilscience for three reasons. Determining limerequirements has been studied extensively andgeneral gUides are available. but the actualmethod of making a recommendation remainscontroversial. Secondly, this was an area ofexpertise with which the authors were relativelyfamiliar. This meant that we did not have tolearn the skills needed for knowledge and ruleextraction at the same time that we were learn­ing to represent the knowledge and rationale inthe trial expert system. As time goes by and wegain more experience. it is becoming clear to usthat this may be an alternative to hiring aknowledge engineer to extract and code theinformation. Finally. although there are impor­tant mathematical calculations in determininglime requirements. there are a large number ofconditions that need to be attached to anumerical recommendation of the lime rate.This is a result of the highly complex soil­plant-elimate-man system whose behavior weare attempting to simulate.

Clancey (1985) has suggested that expertsystems are particularly appropriate for the

solution of problems with several character­istics; he calls these characteristics the"heuristics of knowledge acquisition":

1. Problems amenable to solution withexpert systems

a. Are the solutions to the problemsenumerable? Are there plans. diagnoses. typicalconfigurations or syndromes?

b. Can the problem be subdivided intosmaller classification problems?

c. What is observable? Does causalityplaya role in the problem?

2. Knowledge representation in a rule-basedsystem

a. List all possible solutions to theproblem. Organize these solutions into classes.types. or hierarchies as appropriate.

b. Identify relationships among the data:generalizations. definitions. and qualitativeabstractions. Experts frequently leave outqualitative abstractions. stating associationsinstead in terms of numeric data.

c. Identify and establish heuristics. orrules. that relate data to solutions after thesolutions have been identified.

d. Modeling the experts' thoughtprocesses can be difficult-they frequently tendto use a hypothesis generation-proof procedurethat is difficult to implement with similarflexibility in the expert system.

3. Refinement of an expert system.Experience suggests that one should rapidlydevelop a prototype system and show it to thedomain expert (subject matter expert), carefullynoting his or her reaction (Waterman. 1986).With a sufficiently detailed prototype. thedomain expert can identify with portions of thesystem rationale and will qUickly see flaws andhave the opportunity to suggest modifications.

Constructing an expert system requireslearning concepts of organizing informationand arranging rules and rule components so thatquestions are asked in an efficient and sensiblemanner. These include concepts based on theway the system searches the rules anddetermines what information to ask of the user.The way the system does this is determined bythe structure of the "inference engine."

OUR EXPERIENCE WITH EXSYS

Introduction to EXSYSEXSYSI (Hunington. 1985) is a rule-based

expert system development shell designed to

lUse of brand names does not mean that the authors endorsethis particular software. The discussion would apply to mostrule-based shells.

provide many of the common rule constructionactivities that form the substance of expertsystem development. The shell provides editingfacilities to design output formats. run test datasets. and ensure that modifications have notdisrupted the core logic flow of the system. Theinference engine is mainly backward chainingand provides a switch that will enable the searchto stop after the first valid rule or will cause thesearch to continue until all possible rules areevaluated. Only simple WHY capability isprOVided: it displays the rules that are beingevaluated in the information input mode orprovides the chain of fired rules in support of arecommendation. Programming effort isminimal with this software. There is a loss infleXibility for certain types of expert systemconstruction. This software comes close toproviding an expert system development systemthat computer-acquainted professionals shouldbe able to learn with little effort.

In EXSYS the search procedure followsseveral simple rules. These search rules will bediscussed in the sequence in which they operate.

Choice selection. The first search or "patternmatching" that is done on the knowledge basebegins with the "choices." Choices in EXSYS area list of all the potential conclusions fromwhich the system can choose in presenting finalresults. None. one. or several choices are possi­ble with any consultation or "run" of the system.

During run time. all choices in EXSYS willbe checked by the inference engine to determinewhether they are true. This checking procedureis sequential. The first choice on the list isselected first to determine if it can be proved trueor false. Then the second choice is selected. andso on. The order of the choices is determined bythe person developing the system.

The final results are selected and rated onthe basis of the combination of the probabilitiesassigned to the choices in each of the supportingrules. The inference engine provides probabilityaccumulation in dependent. independent, andaveraging modes. Each choice must have at leastone supporting rule that involves probabilitiesof the choice being correct.

Rule selection. The rule whose THEN partcontains the choice currently being checked willbe selected for analysis. If there is more than onechoice, the rule with the smallest rule numberwill be chosen first; the rule with the largernumber will follow. If two rules have exactly thesame THEN parts (if mathematical evaluation isinvolved. see the relevant section. next page), therelationship between these two condition sets isOR, the same as that between the results of theanalyses on these two rules. But within a rule allconditions in the IF part or in the THEN parthave a relationship AND. For example:

3

~i';;

RUlE 22IF

You have used this system before and you don't need to seedetails of this system .....This condition will be searched

for and, if not found, will bedisplayed for user selection.

THENYou prefer to continue .....This condition matches the

condition in RULE 11 and is whyRULE 22 was selected. (Thisrelationship between rules is alsoknown as chaining.)

THENNo lime is recommended -probability (100/100) .....This is

the first choice in the list ofchoices. This choice caused thisrule to be selected first foranalysis.

RUlE 11IF

You prefer to continueThe soil great group is Paleudults

THENNo lime is recommended-probability (100/100) .....This is

the first choice in the list ofchoices. This choice will causeRULE 11 to be selected first foranalysis. (Explanatory commentadded for this paper.)

In order to test choice 2, the inference enginetries to evaluate (Xl (RULE 20). If, in the wholeknowledge base, only RULES 12 and 15 contain(Xl in the TIiEN part, RULE 12 will be picked firstfor analysis. If both rules are evaluated, [Xl willbe initially assigned as 1 by RULE 12 and reas­signed as 2 by RULE 15. This mechanism enablesmodification of a mathematical value throughseveral rules and different considerations.

Mathematical evaluation. While EXSYSallows numerical comparison in the IF part ofany rule, it also allows the assignment ofnumerical values to mathematical variables inthe TIiEN part of any rule. If more than one rulehas the same variable evaluated in the THENpart, the rule with the smallest number will beevaluated first and the variable will get its firstvalue. The next rule will then be evaluated and,if all of the conditions are satisfied, the variablewill be assigned a new value based on the currentcalculation. For example, we have these rules:

12 IF condition 1TIIEN[X) = 1

15 IF condition 2TI-lEN [X) = [X) + 1

20 IF [X) =2TIiEN choice 2

Summary of EXSYSThe search begins with the selection of a rule

to analyze. The first choice in the list of choicesis selected. A rule that has this chOice in itsTHEN part is selected. If there is more than onerule with the choice in its THEN part. the rulethat has the smallest rule number will beselected.

The IF part of the selected rule is looked at.The first condition is selected and is checked tosee whether it can be proved true or false frominformation already determined to be true. If thecondition cannot be determined to be true orfalse based on information already in thesystem, the system will put the condition onscreen and ask the user which combination ofqualifiers and values is true.

The order of the conditions in the firstselected rule will determine the order of thequestions asked of the user. For example,. onewould prefer that the most likely conditions beasked first. This can be done by placing the mostcommonly selected choice as the first in thechoice list (choice 1). This order should matchthe "directed graph" or decision tree that werecommend be constructed first to document thelogical organization of the expert system. Inother words, the first condition in the ruleshould ask for the most general information.This ensures that the system will reduce the

(search order).....first searched.....second searched

RULE 1IF (qualifier) (value)

fuuprekrro c~tinue

The soil great group is Paleudults

Condition selection. Once a rule is selected,EXSYS proceeds to analyze components of therule- the "conditions." In EXSYS a conditionhas two parts: the "qualifier" and the "value."After selecting which rule to analyze, EXSYSdetermines which condition to analyze. Thisselection is quite logical- the first condition inthe IF part of the rule is selected for evaluation.The purpose of the evaluation is to determine ifthe condition is true or false. The system deter­mines if the condition is true by first searchingthe file of facts or input already concluded to betrue; if the condition is not in the list of facts,the system searches rules that have the condi­tion in their TIiEN parts. If the system finds arule with the condition in the THEN part, itdetermines whether the conditions in the IF partof that rule are true or can be concluded to be truefrom other rules by chaining. If our first condi­tion can be determined to be true, the system willproceed. Otherwise it will, as a last resort, dis­play the first condition in the rule that had ourfirst condition in its THEN part. These condi­tions will be displayed in the form of a qualifierand its values, asking the user to indicate whichones are successive conditions. For example:

4

number of rules to be searched in subsequentsteps. If a very specific condition were placedfirst. it might ask for information that wasirrelevant to most of the system.

The ideal expert system asks the simplequestions first and uses its rules to infer themore difficult. specialized knowledge. Thesystem would be self-defeating if. instead ofasking the simple questions first. it asked thespecialized questions that reqUired expertknowledge. The condition sequence within arule is thus crucial in establishing a searchsequence within the expert system that willdevelop information from the general to thespecific as it progresses through the system.

The ACID4 SystemThe ACID4 knowledge base was developed

from existing information and research experi­ence. Because the priinary objective has been toaddress soil acidity problems in the transmigra­tion area of Indonesia. we have focused onextractable acidity (mostly exchangeable AI) asthe primary cause of yield reduction. Thereference for the knowledge base is a reviewpaper by Kamprath (1984). The main concepts inthe data base are these:

1. Growth-limiting effects are due primarilyto exchangeable AI + H (exchangeable acidity).although. if all cations are present in very smallquantities. some lime is probably needed to pro­vide Ca. It is assumed that toxicity to exchange­able acidity is closely related to AI + H satura­tion so that AI + H saturation is a satisfactorymeasure for diagnosis.

2. Crops vary considerably in their toleranceto exchangeable acidity: extremes are represent­ed by mung bean (very intolerant. tolerating nomore than 0 percent Al saturation) and cassava(very tolerant. tolerating about 75 percent AI + Hsaturation).

3. Organic material seems to reduce limerequirements. The current approximation isthat 10 tons/ha of fresh organic materialreduces the lime requirement by 1 ton/ha.

4. Lime requirements are based on soilanalyses in order to .accurately reflect the soilconditions.

5. Although data are sparse. an attempt ismade to determine the approximate effects oflime quality on the lime requirement. Includedare the neutralization value relative to calciumcarbonate and an estimate of physical reactivityas related to the particle size. The estimate ofneutralization value is a well-defined labora­tory procedure in which an excess of acid isadded to the lime and allowed to fully react. Theexcess acid is back-titrated to determine theunreacted acid for the calculation.

Particle size fractions have been used toestimate the physical reactivity of the lime­stone. Many factors have been studied toestimate the time reqUired after applicationbefore crops can be planted. One of the simplermeasurements of lime quality. as affected byparticle size. is given by the measure of theamount of lime of various particle sizes neededto give approximately equivalent yields. Theequation we used was developed from Figure 4 inBarber (1984):

----------------TonsoITInlerequ~edto----------

Fraction passing attain aooA! Fractiona 60-mesb sieve relative yield increase

----O~90----------------&7---------1~60------

nm w 1~0.55 4.4 1.190.45 5.1 1.380.35 6.2 1.680.25 8.0 2.16

The calculation of lime requirement is basedon the need to neutralize sufficient AI to reducealuminum saturation to the "critical aluminumsaturation" that has been established for thevarious crops (Cochrane et aI.. 1980). Ourmodified form of the equation is:

Lime requirement (t/hal = 1.4(exchangeable acidity- (CAS*ECEC/l00))

where: - exchangeable acidity is the IN KCI extractableAI+H

- CAS is the critical aluminum saturation of the crop- ECEC is the "effective cation exchange capacity"- the value 1.4 represents the relation of the cmol of

CaC03 reqUired to neutralize 1 cmol of AI + H infield studies adjusted for both bulk density anddepth of incorporation. In this case 1.9 cmol of Cawas required for each cmol ofAl + H. the bulkdensity was assumed to be 1.0. and the depth ofincorporation was assumed to be 15 cm.

Preliminary data suggest apprOXimately0.53 cmol KCI-extractable acidity is neutralizedfor each cmol of Ca added as CaC03 (Wade et al..1985). This corresponds to a relationship of 1.9cmol of CaC03 being reqUired for each cmol ofextractable acidity. a value that is similar to theresults reported elsewhere (Kamprath. 1984).This reference points out the need to considerthe effectiveness of lime in neutralizing theextractable acidity. Such data should to beobtained in field studies. if possible. because ofthe need to ensure that one is testing the limingmaterial and soil reactivity under conditionsthat are representative of the situation or for agroup of farms for which the eventual recom­mendation is intended.

Other data and results from the Tropsoilswork in Sitiung. Indonesia. are incorporated.such as minimum requirements of P and K forsoybean. rice. cowpea. and peanut.

5

A directed graph of ACID4 is shown in Figure1. The system is designed to apply to the humidtropics with soils of the Ultisol. Oxisol. andInceptisol orders. The system has additionalinformation pertinent to the Sitiung region. Thegeneral recommendations are based on otherrelationships. such as a general reactivity of 2cmol of CaC03 for each cmol of extractableaCidity. Levels of critical aluminum saturationare. so far. the same for the general recommen­dation as for the specific location in Sitiung.

At present the soil great group isincorporated for the aquic- and fragi- subgroups.We expect to expand this to access a much largerdata base such as that being developed by ourcollaborators for their extensive collection ofsoils data from various surveys and inventories.One possible use of such a data base would be torequest the town or geographic location toobtain a summary of the soil characteristics orspecific problems that might be a problem forcrop production or soil management in the area.

FUTURE APPLICATIONS OF EXPERT SYSTEMSWe believe that knowledge-based systems

offer considerable potential to help us organizeand transmit problem-solving expertise. ThiSshould foster and stimulate application ofagronomic knowledge in concepts rather than assimple facts or observations as in the past.

Expert systems are useful for us in agronomyand soil science and probably in agriculture andbiological science in general for four reasons:

1. Agronomy and soil science deal with ahighly complex. descriptive soil-plant­climate-human system. Usually a large amountof information is necessary to understand andpredict any particular phenomenon- it is notusually possible to reduce this large amount ofinformation to a single rule. theorem. or axiom.Some scientists. however. have succeeded incondensing their information to some extent.although it took them many years and at timestheir entire career. They have developed rules ofthumb or heuristic gUidelines that may notalways work. but usually do. In some cases thisinformation may be all there is. or this may bethe best way to represent the state of knowledgeof the system. It may be that we have in somecases attempted to fit round pegs into squareholes in attempting to attach numbers to suchill-defined phenomena. Certain types of knowl­edge representation- for example. fuzzy systems(Negoita. 1985) and symbolic representation(Chandrasekaran. 1983)- offer a way to repre­sent this approximate information in a moreprecise way than is possible with a singlenumber. Expert systems permit capturing thisexpertise and knowledge so that it need not be

6

learned all over again nor duplicated byfollowing generations.

2. Potential applications of expert systemsin agriculture include potential benefits toresearch. extension. and instruction.

Development of problem-solving protocoland procedure aids in identifying informationgaps. In this fashion. research becomes moreefficient and becomes part of a larger manage­ment strategy. Expert systems can serve as animportant aid in remembering logical segmentsor procedures that eventually can be linked intoa larger rationale. At various stages of develop­ment. expert systems can serve as memory orprocedure aids. "assistants." "associates." orultimately as "experts." depending on the leveland quality of knowledge and skill attained.

A well-designed expert system offers a widerange of possible applications in extension.Repetitive questions by clients might be handledby a system that patiently addresses popularproblems or queries. An incredible amount ofinformation and problem-solving skill could beat the county agent's or extension specialist'sdisposal if even a small fraction of theagricultural expertise were recorded in thisform. New information could be qUickly andaccurately disseminated.

It is clear that agricultural graduates must beequipped to manage. evaluate, and developinformation to a greater extent than ever before.Acquaintance and skill with knowledge-basedsystems would permit improved problem­solVing ability as one additional informationmanagement tool. Concepts of machine learningand intelligence stimulate thought and reflec­tion on improving human skills in this area.Inclusion of knowledge-engineering concepts ininstruction and graduate programs promotesawareness of information technology skills andtheir value in managing information.

3. If expert systems technology were Widelyimplemented in agriculture, we should see arapid advance in agricultural science from anessentially phenomenological stage to one inwhich the knowledge is more highly structuredand organized. This should lead to a moreadvanced stage of development of the science,which, in tum, should lead to more emphasis onprinciples. Developments in other areas ofscience suggest this would pave the way for moretheory development with a consequent increasein the number and significance of researchbreakthroughs.

4. From the nature of fifth generationresearch it is apparent that expert systems arethe first of many innovations that we can expectfrom the application of microprocessors toinformation technology development. In manyrespects we have been using microcomputers to

Begin:Consult ACID4soil database, , ,

Begin DisplayQuit

asking questions this graph

I, ,Ultisol; Oxisol; Histosol; Vertisol; No information

Inceptisol Alfisol; Entisol; END:

Spodosol; Mollisol;Aridisol

Soil great groupis:

t t

Known Unknown,Give great group

Crop is:Critical AI sat.Dry Rice; Peanut;

constraints Maize; Soybean; for the crop

Cassava; Mungbean; tPasture grasses;Pasture legumes Effective CEC:

tDepth of

Soil Bulk Density: Extractable AI+Hincorporation:

Adjust lime rate Soil pH: Biocarbonate P =for quality and (Consistent

fineness: with AI data?) ,Exchangeable K =

Recommendations: Cautions: Notes:1. Amount of lime 1. Data consistent? 1. Subsoil AI?2. Amount of P 2. Adequate depth? 2. AI tolerance?3. Amount of K 3. Too much lime? 3. Soil constraint?

IEND Perform

Change input partial budgetand rerun? economic analysis

Figure 1. Directed graph of ACID4 expert system.

7

do the same things that we have been doing formany years- only faster and more conveniently.The maturing of this technology will lead toconceptual innovations. Almost certain to comesoon will be powerful reasoning and learningcapability beyond that currently provided bylanguages such as PROWG. Agriculture shouldbe in a position to benefit much more from suchtechnology than the already highly structuredsciences such as physics and chemistIy.

LITERATIJRE CITEDACM. 1985. Special section on architectures for

knowledge-based systems. I. The role offrame-based representation in reasoning. RFikes and T. Kehler: II. Rule-based systems. F.Hayes-Roth: III. Logic programming. M. RGenesereth and M. L. Ginsberg. Communica­tions of the ACM. September 1985. 28:903­941.

Barber. S. A. 1984. Liming materials andpractices. In F. Adams (ed.). Soil aCidity andliming. 2nd ed.• Amer. Soc. Agron.. Madison.Wisconsin.

Barr. A. and E. A Feigenbaum. 1981. The hand­book of artificial intelligence. Kaufmann. LosAltos. California.

Chandrasekaran. B. 1983. Expert systems:Matching techniques to tasks. Presentation atthe New York University Symposium onArtificial Intelligence Applications forBusiness. May 18-20. 1983.

Clancey. W. R 1983. The epistemology of rule­based expert systems- a framework forexplanation. Artif. Intell. 20:215-251.

Clancey. W. R 1985. Knowledge acquisition forclassification expert systems. ProceedingsACM 1984 Annual Conference: The FifthGeneration Challenge. San Francisco.California.

Clocksin. W. F.. and C. S. Mellish. 198!.Programming in PROLOG. Springer-Verlag.New York.

8

Cochrane. T. T.. J. G. Salinas. and P. A. Sanchez.1980. An equation for liming acid mineralsoils to compensate crop aluminum tolerance.Trop. Agr. 57: 133-140.

Hayes-Roth. F.• D. A. Waterman. and D. B. Lenat.1983. Building expert systems. Addison­

Wesley Pub. Co.• Reading. Massachusetts.HUnington. D. 1985. EXSYS. version 3.0. Expert

System Development. Albuquerque. NewMexico.

Kadane. J. B.. J. M. Dickey. R L. Winkler. W. S.Smith. and S. C. Peters. 1980. Interactiveelicitation of opinion for a normal linearmodel. J. Amer. Statis. Assoc. 75:845-854.

Kamprath. E. J. 1984. Liming acid soils of thehumid tropics. In F. Adams (ed.), Soil aCidityand aiming. 2nd ed.. Amer. Soc. Agron..Madison. Wisconsin.

Michaelson. R H.• D. Michie. and A Boulanger.1985. The technology of expert systems. BYTEMagaZine. April 1985.

Michalski. R S .• and R L. Chilausky. 1980.Learning by being told and learning fromexamples: An experimental comparison of thetwo methods of knowledge acquisition in thecontext of developing an expert system forsoybean disease diagnosis. Intern. J. PolicyAnal. and Info. Systems 4: 125-161.

Michalski. R S.. J. H. Davis. V. S. Bicht. and J. RSinclair. 1982. PLANT/ds: An expert systemconsulting system for the diagnosis of soy­bean diseases. Proceedings 1982 EuropeanConference on Artificial Intelligence. Pp.133-138.

Negoita. C. V. 1985. Expert systems and fuzzysystems. Benjamin Cummings Pub.. MenloPark. California.

Wade. M. K.. Heryadi. Aljabri. F. Agus. and E.Joniarta. 1985. Liming in transmigrationareas. Report submitted for the TropsoilsAnnual Report. 1985. Tropsoils/Indonesia.

Waterman. D. A 1986. A gUide to expert systems.Addison-Wesley Pub. Co.. New York.

DISCLAIMERReference to a company or product name does not imply approval or recommendation of the product bythe College of Tropical Agriculture and Human Resources, University of Hawaii, or the United States Depart­ment of Agriculture to the exclusion of others that may be suitable.

Hawaii Agricultural Experiment StationHITAHR, College of Tropical Agriculture and Human Resources, University of Hawaii at ManoaNoel P. Kefford, Director and Dean

RESEARCH EXTENSION SERIES 089-03.88 (2M)