---
title: "Empirical and Model-based Reasoning in Expert Systems"
authors:
  - "Phyllis Koton"
doi: ""
url: "https://ijcai.org/proceedings/1985/56"
pdf: "https://ijcai.org/proceedings/1985/0056.pdf"
venue: "Proceedings of the Ninth International Joint Conference on Artificial Intelligence"
year: 1985
pages: "297-298"
track: "Expert Systems"
keywords:
---

# Empirical and Model-based Reasoning in Expert Systems

Phyllis Koton

*Proceedings of the Ninth International Joint Conference on Artificial Intelligence*, pages 297-298.

## Abstract

Many expert systems are now being written which rely on highly-compiled, empirical knowledge for their reasoning power. Model-based reasoning has significant theoretical advantages. I constructed two expert systems for the same domain , one using large-grained compiled knowledge, the second using model-based reasoning. The use of model-based reasoning resulted in improved knowledge accessibility and flexibility, and expanded problem-solving ability. Many current expert systems rely on highly-compiled, large-grained knowledge (heuristics, empirical associations, "r ules of thumb") for their reasoning power.* An alternative approach, model-based reasoning, uses a detailed model of the objects in the domain and the operations that act on those objects. The relative problem solving abilities of these two different types of reasoning have not previously been experimentally compared. I constructed two programs that solve problems in the same domain, with the same objectives, but each using a different reasoning method. GENEX (I{oton , 1983) and GENEX II (Koton, 1985) solve problems about the behavior of bacterial operons, ** a subfield of molecular biology.

## Full text

## EMPIRICAL AND MOD8L-BASED REASONING IN EXPERT SYSTEMS

P. A. Koton

Laboratory for Computer Science Massachusetts Institute of Technology Cambridge, Massachusetts 02139

grams take as input a description of an operon (real or imaginary) and information on whether or not. t.he genes of that operon arc expressed, then attempt to deduce the biological control mechanism causing the observed behav­ ior of the operon. GENEX uses large-grained empirical knowledge to solve problems, and GENEX II uses model­ based reasoning. GENEX II can solve a greater variety of problems and more difficult problems than G ENEX.

The expert knowledge in the GENEX program consists of 62 pieces of knowledge relating observed phenomena to possible causes. Each piece of information is highly com­ piled and, in general, incorporates several inference steps. For exan1ple:

IF the mutation is in the promoter AND the gene is not expressed THEN possibly RNA polymerase cannot bind to the promoter.

G ENEX performs well on many problems. However, the types of problems which it. can be used to solv<! is limited. It can only be guaranteed to yield a correct solution for sim­ ple problems involving a prototypical operon containing at most one mutation. For some more difficult problems, in­ volving more than one mutation, GENEX gives the correct answer, but for others, it gives an incorrect solution. This limitation may be attributed to the program's reliance on reasoning via empirical associations.

In order to overcome the limitations of GENEX, GENEX II

was implemented. This program contains a description of each object and mechanism involved in operon control. For instance,

the BINDING-SITE of RNA-POLYMERASE

i.s the PROMOTER.

a MOLECULE(M) BINDS to a SITE(S) if S is the BINDING-SITE for M and M and S have COMPLEMENTARY-CONFORMATIONS and nothing else is BOUND to S and M is FREE to BIND.

## ABSTRACT

Many expert systems are now being written which rely on highly-compiled, empirical knowledge for their reason­ ing power. Model-based reasoning has significant theo­ retical advantages. I constructed two expert systems for the same domain, one using large-grained compiled knowl­ edge, the second using model-based reasoning. The use of model-based reasoning resulted in improved knowledge accessibility and flexibility, and expanded problem-solving ability.

Many current expert systems rely on highly-compiled, large-grained knowledge (heuristics, empirical associations, "rules of thumb") for their reasoning power.* An alter­ native approach, model-based reasoning, uses a detailed model of the objects in the domain and the operations that act on those objects. The relative problem solving abilities of these two different types of reasoning have not previously been experimentally compared. I constructed two programs that solve problems in the same domain, with the same objectives, but each using a different rea­ soning method. GENEX (I{oton , 1983) and GENEX II (Koton, 1985) solve problems about the behavior of bacte­ rial operons, ** a subfield of molecular biology. Both prothe EXPRESSION of OPERON(O) INITIATES if PART-OF(O) is P and P is a PROMOTER and RNA-POLYMERASE BINDS to P.

This r<'sParch was supported in part by the National Institutes of Hf'alth Graut. No. 1 POl LM 03374-04 from the National Library of M<·diriiH\ and in part by National Imtitutf's of Hmlth Grant No. 1 P41 RR 010!l6-05 from the Division of Research R<·sources. * Tlw~e concepts arf' not. strictly logically <'quivalent., hut in cnr­

r<'nt practice syst<'ms based on empirical or heuristic associations typically do express their knowledge iu large-graiiu·rl, compiled form.

"'Th<' term ntJeron r<'fns to a genetic mnt.rol syst.<'lll in which thl' activity of a sd of .·tructuml gene.·, ro<ling for nll'tabolically r<'lated proteins r<>quired by the cell, is governed hy the inter­ action of a regulatory protein with an adjacent region of DNA (called the operator). In order for the operon proteins to be made, the enzyme RNA polymeras!' mmt be able to bind to the DNA at a specific site, known as the Jlromoter. It can only do this if the operator, which overlaps the promoter, is not blockro by the presence of the bound regulatory protein. Whether or not the regulatory protein hinds at the operator depends on whether the cdl needs to make more of the op<'ron proteins.

an OPERON(O) is EXPRESSED if 0 is INITIATED and 0 is not ATTENUATED and 0 is TRANSCRIBED and 0 is TRANSLATED.

This knowledge base, written in Prolog, contains just over 250 clauses. The descriptions include the different parts of the operon, the role that each part plays in the functioning of the operon, how the parts interact, and how lower-level mechanisms combine into higher-level behav­ ior. The knowledge base for GENEX II contains essentially the same information that GENEX has, but the informa­ tion that is "compiled into" GENEX is stated explicitly in GENEX II. These descriptions enable the program to reach its solutions by constructing a model of the operon and simulating the processes that act on it rather than doing a match between the observed phenomena and the knowledge in the system, as GENEX does.

GENEX II can deal with two large classes of problems which GENEX was unable to solve: the combined effect of multiple mutations on gene expression, and problems involving non-prototypical operons. An example of the type of problem that GENEX II can solve and on which GENEX fails is shown below.*

Bacterium pednntia produces a protein, Sporulin, that makes other bacteria sporulate. Sporulin syn­ thesis is inducible by chalk. In the absence of chalk a repressor produced by gene spaR prevents expres­ sion of the Sporulin structural gene spoS.

A variation carrying two mutations can be made in B. pedantia by recombination. Most double mu­ tants that carry both a constitutive spoR and a constitutive spoO mutation [mutations in the regula­ tory gene and the operator of the spo gene that cause its product to be made continually] are constitutive, as ex­ pected. However, one particular double carrying mutations spoR43 and spo097 is anomalous: it pro­ duces Sporulin only when chalk is absent. spoR43 is known to be missense [a missense mutation is one that changes the structure of the resulting protein]. Explain briefly how the repressor and operator are func­ tioning in the double mutant.

*This problem is from an MIT undergraduate genetics exam. It has been simplified for this paper. The full text of the problem can be found in (I<oton, l!l85) .

GENEX II solves this problem by first noting that the mutations have introduced uncertainty about the three-di­ mensional shape of objects in the domain. Complementar­ ity of shape partially determines which objects can bind to which others. The program generates all possible pairs of shapes (in the abstract sense; i.e. to which other objects is this object's shape complementary) which are consistent with the mutations given and the observed behavior of the operon. One possibility it generates is that the regulatory protein has been mutated in such a way that only when it is bound to chalk can it bind to the operator. GENEX II simulates the behavior of the operon under this set of circumstances. It reasons along the following lines:**

In order for the operon to be expressed, RNA poly­ merase must be able to bind to the promoter. It can­ not bind to the promoter if something is already bound there. Because the operator overlaps the promoter, any­ thing bound to the region where the operator and promoter overlap is bound to the promoter. The system then looks to see if anything can be bound to the operator. The hy­ pothesis that only when chalk is bound to the mutated regulatory protein it is in the proper conformation to bind to the mutated operator, results in the system reaching the conclusion that in the presence of chalk, RNA polymerase cannot bind and the gene product is not made. This is consistent with the observed behavior of the operon, and will be reported as a possible solution.

When the same problem is given to GENEX, it recog­ nizes that the spo gene is being repressed by chalk, and thus when chalk is present the gene product is not made. GENEX can not discover the mechanism by which this occurs, because this particular combination of mutations does not correctly match any of its heuristics. It partially matches two heuristics, one stating that the mutated op­ erator cannot bind a normal regulatory protein, and the other stating that the mutated regulatory protein cannot bind to a normal operator, and can go no further.

Certainly, any given problem which is currently un­ solvable by GENEX, or other systems which use compiled knowledge, could be solved by giving the program a new piece (or pieces) of information. However, this would im­ prove the performance of the system only for that one type of problem. Compiled knowledge captures domain knowl­ edge as is; it does not capture the underlying model. Be­ cause such programs solve problems by matching the cur­ rent situation against a set of predetermined situations, the knowledge base anticipates situations that may arise. For small domains, it may be possible to enumerate every pos­ sible situation and encode a response for each one.*** In a large domain, it may be extremely difficult to predict every possible state of the system. In contrast, a sys tem which uses model-based reasoning, such as GENEX II , can generate new situations based on its underlying model of the domain.

**The program's complete solution is 13 pages long. Interested readers are referred to (I<oton, 1985).

***This technique was used successfully in PUFF (I<unz, l!l78) .

Empirical "rules of thumb" often contain implicit pre­ conditions. For example, one such rule in GENEX states that "deleting the operator results in increased expres­ sion of the operon." The implicit precondition here (as it clearly must be in similar heuristics for other programs) is "all other things being equal." If all other things are not equal, the validity of the conclusion is uncertain. Im­ plicit preconditions can make it difficult, if not impossible, to determine the conditions for which a "rule of thumb" can be applied correctly. However 1 explicitly stating every precondition may be tedious - consider having to explic­ itly encode "all other things being equal" in a system with hundreds of components.

Another limitation of compiled knowledge is that the domain knowledge is contained implicitly, and thus is in­ accessible. For instance, the GENEX rule shown above en­ codes the rule of thumb that a mutation in the promoter and a decrease in transcription of the operon suggests that RNA polymerase cannot bind to the promoter. If given a problem that included a promoter mutation and a de­ crease in gene product, GENEX applies this rule and sug­ gests that the promoter mutation caused the decrease in gene product. However, GENEX has no knowledge of the mechanism which makes this true, i.e. that the promoter is where the RNA polymerase binds, if the promoter is deleted the polymerase cannot bind, and if the polymerase cannot bind the operon cannot be transcribed. This infor­ mation is explicitly encoded in GENEX II, as is illustrated above. The advantages of such increased accessibility have been extensively discussed in (Swartout, 1981).

Knowledge encoded implicitly is also inflexible. One piece of knowledge may be contained in any number of places in the system, so in order to change that knowledge it must be changed every place it is used.

When model-based reasoning is used, the program cre­ ates its solution anew for each problem, based on its model of the domain, rather than relying on pre-compiled solu­ tions. Theor.etically, the program should be able to solve any problem that is derivable from the model of the do­ main.

As always, there are tradeoffs. Model-based systems such as GENEX II require a more complicated control struc­ ture to reduce the amount of search for possible solutions. The finely-grained knowledge in GENEX II creates infer­ ence chains that are about four times longer than those generated by GENEX in the solution of the same problem. This results in a slower-executing program. Perhaps the ideal system would use both empirical associations for speed and model-based reasoning for improved problem­ solving ability and better understanding of its reasoning, as described by (Barnett, 1982).

## Acknowledgments

I would like to thank Peter Szolovits, who supervised the research on wllich this paper is based, and Ramesh Patil, who introduced me to many of the issues raised in this paper.

## References

1. Barnett, J. "Some Issues of Control in Expert Sys- tems." In Proc. of the International Conference on Cybernetics and Society. Seattle, Washington, Octo- ber, 1982, pp. 1-5.
2. Koton, P. "A System to Aid in the Solution of Prob- lems in Molecular Genetics." SM Thesis, MIT Lab- oratory for Computer Science, May 1983.
3. Koton, P. "Towards a Problem Solving System for Molecular Genetics," Technical Report 338, MIT Lab- oratory for Computer Science, Cambridge, Massachusetts, 1985.
4. Kunz, J . "A Physiological Rule-based System for In- terpreting Pulmonary Function Test Results," Heuris- tic Programming Project Report HPP-78-19, Stan- ford University Computer Science Dept., Stanford, California, 1978.
5. Swartout, William R. "Producing Explanations and Justifications of Expert Consulting Programs," Tech- nical Report 251, MIT Laboratory for Computer Sci- ence, Cambridge, Massachusetts, 1981.