![]() | ME290M
Spring 1999, T-Th 12:30-2:00 pm
|
| [ Home | Info | Syllabus | Readings | Students | Homework | Resources | News | Chat] |
1 Nonmonotonic logic
Nonmonotonic logic is concerned with reasoning that allows making assumptions that are not guaranteed to be correct (and thus involves a "nonsound" logical procedure), but are at least consistent with the statements in the knowledge base.
In order to solve everyday problems, we often resort to common-sense or default reasoning when the information required for a decision is incomplete or too complex to easily process. When new or more complete information is added, we may change our assumptions and make different decisions. Much of human learning is of this nonmonotonic nature. We create certain models of the world as children. As we mature and learn from experience, we change some parts of those models.
If expert systems are to emulate human experts, can they be made to "learn" and update their models of the world when contradictory information is provided to them? If we allow nonmonotonic reasoning in an expert system, it will be necessary to
(1) Check each new statement with consistency of other statements in the knowledge base.
(2) If there is an inconsistency, some means of dealing with it must be incorporated in the expert system control structure.
(3) When one statement is deleted from the knowledge base, the control system must go back over other statements whose proofs depend on the deleted statement and either eliminate them or find new proofs with the new knowledge base.
Nonmonotonic reasoning can sometimes be avoided by adding conditionals to a knowledge base of classical logic. Consider the classic AI example: All birds fly, or in prefix predicate calculus:
(IF (BIRDS x)(FLY x))
Although most birds do fly and this statement may be true most of the time, there are exceptions. What about ostriches, penguins, kiwis, or dead birds? We could revise the logical sentence and still retain monotonicity:
(IF(AND ((BIRD x) (NOT (OSTRICH x)) (NOT (PENGUIN x))(NOT (KIWI x))
(NOT (DEAD x)))(FLY x))
The problem with this logical sentence is that it is long and one must know ahead of time whether x is an ostrich, penguin, kiwi, or not dead. What if this information is not available? What if all that we know is that x is a bird? Can x fly?
1.1 Qualification Problem
Even though there are many qualifications in the logical sentence in the last section, we can think of many more: Does the bird have wings? Does the bird have feathers? Is it a baby bird? Is it an injured bird? and so on. This illustrates what is called the qualification problem. In order to represent universally quantified logical sentences accurately for every possible case in the world, a seemingly infinite number of qualifications may be required.
1.2 Default Rules
An alternative approach is the use of default rules. Like other rules of inference, a default rule consists of a set of premises and a conclusion. The logical sentence that is shown to be true is true, unless there is evidence to the contrary. A default rule is represented by a double line to distinguish them from regular rules of inference. The rule below states that whenever a knowledge base contains a logical sentence of the form a, it is legal to add ß to the knowledge base so long as it is consistent with the knowledge base. (By consistent, I mean that there is no rule that violates it).
a
======
b
Some systems will collect possible exceptions associated with a default rule as an explicit part of the rule. Consider the "birds fly" example below:
(BIRD x)
======== D^ R1 R2 . . . Rn
(FLIES x)
The logical sentence above is true, if and only if it does not violate the logical sentences R1 R2 . . . Rn. Where
R1: (IF (OSTRICH x) (NOT (FLIES x)))
R2: (IF (PENGUIN x) (NOT (FLIES x)))
R3: (IF (KIWI x) (NOT (FLIES x)))
R4: (IF (DEAD x) (NOT (FLIES x)))
Note that this differs from the original example (IF (BIRD x) (FLIES x)). It is weaker in that it doesn't allow one to conclude that all birds fly, only birds that are not known not to fly. It is stronger than the lengthy qualified "AND" statement in that it allows one to conclude that a particular bird flies even if there is no information about whether the bird is dead, a penguin, a kiwi, or an ostrich.
1.3 Example of Default Reasoning
Given the above logical sentences, suppose I am also given the following:
1. (IF (MAGPIE x) (BIRD x))
2. (MAGPIE HECKLE)
Thus
(BIRD x)
========= D^ R1 R2 . . . Rn
(FLIES x)
implies that (FLIES HECKLE). If I add (DEAD HECKLE) to the knowledge base, this violates (FLIES HECKLE) and deduction stops -- there is a contradiction. The default rule can not be applied.
Nonmonotonic Algorithm
1. Try to prove (NOT b)
2. If not provable use (IF a b)
Note: the test for consistency must be implemented with caution. As discussed in Chapter 6, provability is usually only semi-decidable. In a semi-decidable system, when a logical sentence is not provable, an inference procedure may run forever. Since consistency of a logical sentence is equivalent to nonprovability of its negation, this means that the consistency test can run forever.
1.4 Conflicting Default Rules
Unfortunately, if we allow nonmonotonic logic, some default rules may conflict with other default rules. For example consider the default rules below: The problem is deciding what to conclude about the color of Eli, the albino elephant.
(ELEPHANT x) (ALBINO x)
============== ===============
(COLOR x GRAY) (COLOR x WHITE)
One way to handle this dilemma is to order all of the default rules. Generally, the more specific the default, the higher the priority. Thus in this case, there are fewer white elephants than gray elephants, thus the albino rule is of higher priority. There are fewer penguins (ostriches or kiwis) than birds; hence a default rule asserting that penguins (ostriches or kiwis) don't fly should be of higher priority than one that says that all birds fly.
1.5 Truth Maintenance System (TMS)
Examples of implemented nonmonotonic logic system are Doyle's "Truth Maintenance System (TMS)" and de Kleer's "Assumption-Based Truth Maintenance System (ATMS)". They are designed as a subsystem for insuring truth maintenance in other reasoning programs. They check for inconsistencies and initiate their own reasoning mechanisms to resolve the inconsistency and appropriately modify the knowledge-base. More detail on TMS and ATMS can be found in the following references:
Artificial Intelligence, April, 1980 (special issue on nonmonotonic reasoning)
Doyle, J., "A Glimpse of Truth Maintenance," in Artificial Intelligence: An MIT Perspective, Vol. I, MIT Press, Cambridge, Mass, 1979 (Math Library).
Doyle, J., "A Truth Maintenance System," Artificial Intelligence, Vol. 12, No. 3, 1979.
de Kleer, J., "An Assumption-Based Truth Maintenance System", Artificial Intelligence, vol. 28, No. 2, 1986.
Rich, E. , Artificial Intelligence, McGraw Hill, pp. 181-184.
1.6 The Closed World Assumption
As discussed earlier, a theory ¡ is complete if it is possible to derive either P or (NOT P) from the database for every logical sentence P that is implied from the database. If a system is not closed, one method for augmenting a theory for closure is to complete it using the closed-world assumption (CWA). The CWA completes a theory by adding the negation of any atomic logical sentence to the theory whenever the atomic logical sentence does not logically follow from the original database D [Reiter 1978]. In other words when using the CWA, we assume that a logical sentence is not true if it is not implied. The CWA is nonmonotonic because this set of augmented beliefs will shrink as we add new logical sentences to D.
1.7 Objections to Nonmonotonic Inference
Formal inference systems based on nonmonotonic logic are in general non-semidecidable. The reason is that in order to apply a default rule, the system must check for contradictions in the default conditions and consistency of the default conclusions. Thus in addition to proving that the default conditions are satisfied, it must also try to prove the negation of the conclusion to verify consistency of these conclusions. Since any first-order logic system is semi-decidable, at best, this means that it can only be guaranteed to stop in finite time if the negation is true, but might run forever if the conclusion is consistent, i.e., the negation of the conclusion can not be proven true.
Thus nonmonotonic systems suffer from the following two major problems. First, nonmonotonic systems can be excruciating slow because of the repeated need for consistency checking. Secondly, the non-decidability of nonmonotonic inference is such that any system using this logic will necessarily make mistakes from time to time (or loop indefinitely). See [Israel, 1980] or [Perlis, 1986] for more discussion on this topic.
2. Three-Valued Logic
Classical logic allows just two truth values; True or False (nil). Kleene's [1952] 3-valued logic was originally designed to accommodate undecided mathematical statements. The third truth value represents an "undecided" value ("u") to indicate a state of partial knowledge. If a component of a logical sentence is not known, its truth value is designated as "u". Three-valued logic provides the mechanism to draw conclusions when the information implies a conclusion and propagates lack of information, otherwise. For example the conjunction (AND P Q) would be assigned the value of "True" if both P and Q are true, "False" if either P or Q are false, and "Undecided" if neither P nor Q are false but either P or Q are undecided. A complete set of truth tables are provided below.
Other versions of 3-valued logic can be found in the following references.
Haack, S., Deviant Logic, Cambridge Univ. Press, 1974.
Haack, S., Philosophy of Logics, Cambridge Univ. Press, 1974.
Kleene, S., Introduction of Metamathematics, Van Nostrand, 1952.
Lukasiewicz, J, "On 3-valued Logic," in: McCall, S., Polish Logic, Oxford Univ. Press, 1967.
Lukasiewicz, J, "Many-Valued Systems of Propositional Logic," in: McCall, S., Polish Logic, Oxford Univ. Press, 1967.

Fig. 1. Three Valued Logic Table
3. Fuzzy Logic and Commonsense Reasoning (Lotfi Zadeh's papers to be put on reserve in the Engineering Library for Week 11 lecture.)
Fuzzy logic goes beyond 3-valued logic in generalizing classical logic to multiple values. In fuzzy logic the set of truth values can fall anywhere in the interval [0,1]. Zadeh advocates the use of countable subsets of these values to make the complexity manageable. He refers to these structured subsets as linguistic truth values such as {true, mostly true, more or less true, not very true, false}.
Let me provide you with a quick introduction to fuzzy sets via a simple example. The main idea in fuzzy set theory is that an element has a degree of membership in a fuzzy set. Consider the fuzzy set "thinness" for a pressure bearing cylinder. The elements are the ratios of radius to thickness of pressure bearing cylinders and their degree of membership depends on their numerical measure of the ratios as shown in the table below:
|
radius/thickness |
"thinness" Degree of Membership |
|
r/t ² 4 |
0.0 |
|
4< r/t ² 6 |
0.1 |
|
6< r/t ² 7 |
0.2 |
|
7< r/t ² 8 |
0.4 |
|
8< r/t ² 9 |
0.8 |
|
9< r/t ² 12 |
0.9 |
|
³ 12 |
1.0 |
The multivalued truth values (membership function "
m" ) in fuzzy logic are calculated as follows:m(NOT P) = 1 - m (P)
m (AND P Q) = Min ( m (P), m (Q) )
m (OR P Q) = Max ( m (P), m (Q) )
m (IF P Q) = Max ([1 - m (P)], m (Q) )
In the 1950s the English economist Shackle [1961] proposed a nonprobabilistic framework for decision making based on the notion of "possibilities". In the 1960s Dempster [1967] introduced the notions of upper and lower probabilities, which were not additive, for dealing with incomplete data. Shafer built on the Dempster model and interpreted these upper and lower probabilities as degrees of plausibility and belief, respectively and introduced the approach to uncertainty management sometimes referred to as "Dempster-Shafer" theory. One distinction of this theory is that it does not require a prior probability for each individual event, but a prior measure of "belief" to classes of events.
"Possibility theory" was introduced by Zadeh in 1977 based on his previous notions of fuzzy sets developed in the 1960s. In possibility theory, the uncertainty of an event is described both by the degree of possibility of the event itself and by the degree of possibility of the contrary event, with some weak relationship between the two. The complement (normalized to 1) of the possibility of the contrary event can be interpreted as the degree of necessity (certainty) of the event itself [Dubois and Prade, 1988].
4. Probabilistic Inference
Unlike a network of computer data bases, human knowledge is dispersed and autonomous. Virtually everyone has information relevant to solving a problem if only this information can be tapped. Incorporating this human knowledge in computer systems is the goal of knowledge engineering. When problems involve failures and uncertain events, as in engineering diagnostic problems, the knowledge is seldom of a binary nature and thus classical approaches to expert systems may be inadequate. Multivariate logics using probability measures may be needed. I refer to the discipline devoted to representing and encoding probabilistic knowledge as probabilistic knowledge engineering. Probabilistic inference will be covered in the next weeks.
References
Brachman, R.J., " 'I Lied about the Trees' Or, Defaults and Definitions in Knowledge Representation, ", The AI Magazine, Fall 1985, pp. 80-93.
Dempster, A.P., "Upper and Lower Probabilities Induced by a Multivalued Mapping, Ann. Math. Stat., Vol. 38, pp. 325-339, 1967.
Dubois, D. and H. Prade, Possibility Theory: An Approach to Computerized Processing of Uncertainty, (trans. by E.F. Harding), Plenum Press, 1988.
Genesereth, M.R. and N.J. Nilsson, Logical Foundations of Artificial Intelligence, Morgan Kaufmann Publishers, 1987.
Ginsberg, Matthew L. (ed.), Readings in Nonmonotonic Reasoning, Morgan Kaufmann Publishers, Inc., Los Altos, CA, 1987.
Israel, D.J., "Whats Wrong with Non-monotonic Logic?", in Proceedings of the First Annual National Conference on Artificial Intelligence, pp. 99-101, 1980. (Also in Ginsberg, 1987, pp. 53-55.)
McCarthy, J., "Circumscription - A Form of Non-monotonic Reasoning," Artificial Intelligence, Vol. 13 (1-2), pp. 27-39, 1980. (Also in Ginsberg, 1987, pp. 145-151.)
McCarthy, J., "Applications of Circumscription to Formalizing Commonsense Knowledge," Artificial Intelligence, Vol. 28 (1), pp. 89-116, 1986.
Perlis, D., "On the Consistency of Commonsense Reasoning," Computational Intelligence, Vol. 2, pp. 180-190, 1986. (Also in Ginsberg, 1987, pp. 56-66.)
Reiter, R., "On Closed World Data Bases," in Gallaire, H. and Minker, J. (eds.), Logic and Data Bases, New York: Plenum Press, 1978, pp. 55-76. (Also in Ginsberg, 1987, pp. 300-311.)
Reiter, R, "A Logic for Default Reasoning," from Artificial Intelligence, vol. 13, no. 1-2, pp. 235-249, 1980. (Also in Ginsberg, 1987, pp. 68-93.)
Rich, E., Artificial Intelligence, McGraw-Hill Publishing Co., 1983.
Shafer, G, "Non-additive Probabilities in the Works of Bernoulli and Lambert, Arch. Hist. Exact. Sci., vol. 19, pp. 309-370, 1978.
Shackle, G.L. S., Decision, Order and Time in Human Affairs, Cambridge University Press, Cambridge, 2nd Edition, 1961.
Turner, R., Logics for Artificial Intelligence, Ellis Horwood Limited, U.K., 1985.
Zadeh, L.A., "Outline of a New Approach to the Analysis of Complex Systems and Decision Processes," IEEE Trans. Syst. Man Cybern., Vol. 3, pp. 28-44, 1973.
Zadeh, L. A., "The Role of Fuzzy Logic in the Management of Uncertainty in Expert Systems, Approximate Reasoning in Expert Systems, Elsevier Science Publishers B.V. (North-Holland), 1985, pp. 3-31.
| [ Home | Info | Syllabus | Readings | Students | Homework | Resources | News | Chat] |
Last updated: 2 March 99