You are on page 1of 8

Deductive Reasoning Systems

Wolfgang Bibel, Technische Universität Darmstadt, Darmstadt, Germany


Christoph Kreitz, Institut für Informatik, Universität Potsdam, Potsdam, Germany
Ó 2015 Elsevier Ltd. All rights reserved.
This article is a revision of the previous edition article by W. Bibel, volume 5, pp. 3346–3349, Ó 2001 Elsevier Ltd.

Abstract

This article explains the relevance of reasoning for prediction and explanation, i.e., for abilities that are fundamental for
human survival. Since humans are prone to mistakes in their reasoning, reasoning systems offer desirable support. Deductive
reasoning systems operate on a model of reasoning that results from a number of abstractions applied to human reasoning.
These abstractions concern the language, the inferential relationship, and the syntactic nature of inference. By way of these
abstractions, other modes of reasoning such as inductive reasoning are also covered. Basic issues concerning research in
deductive systems, their performance and applications, and the psychology of human reasoning, are summarized.

Reasoning is one of the fundamental capabilities associated with embryonic state. Therefore, the study of reasoning so far can
intelligence as manifested in humans. It is a mental activity by only be carried out at a higher level of abstraction. For more
which we generate knowledge not available prior to the act of than two millennia, this study has been carried out at the level
generation. Experience shows that knowledge generated by of language, thus abstracting from the processes that engage
reasoning on the basis of already available knowledge enables many subsystems in the brain including, for instance, those
humans involved in the mental imagery system (Kosslyn, 1995). Since
natural language is still too complex as a basis for the study
l To predict states of the world not experienced before and to
of reasoning, further abstractions have been used. These are
act accordingly and
discussed in the present section.
l To explain phenomena experienced in the world so that we
can understand the reasons why they happened, and make
them happen again or avoid them in the future.
Formal Languages
Prediction and explanation, and their combination, are
Any spoken or written text in natural language involves
fundamental and characteristic ingredients of the human way
numerous linguistic constructs that carry a variety of informa-
of life. In consequence, this is the case for reasoning as well,
tional content. Since reasoning relates assertions, we may for
thus demonstrating its fundamental importance.
present purposes abstract from most of these linguistic
Experience also shows that humans have a sense for the
constructs and restrict our focus on propositions (in the sense
correctness of reasoning chains and also an understanding that
of logic). “The sun is shining” is a simple example of such
correct reasoning chains do not result in arbitrary predictions or
a propositional statement.
arbitrary explanations. Rather, there is a correlation between
Because of the complexity, ambiguity, and redundancy of
the correctness of the underlying chain of reasoning and its
natural language constructs even after the restriction just made,
usefulness for predictions or explanations in the real world:
we further abstract from such propositions by use of some
only predictions and explanations based on correct reasoning
formal language. This should be rich enough to be able to
chains turn out to be useful in general.
express all our propositional knowledge, but simple enough to
On the other hand, humans are prone to mistakes in their
be useful for the manipulations needed in modeling reasoning,
reasoning. In coping with this and other weaknesses lies the
and concise enough to avoid ambiguities and redundancies.
motivation for the mechanization of reasoning in the form of
There are a number of formal languages meeting these criteria.
reasoning systems. They might assist our reasoning, point out
They differ in the level of abstraction taken for the represen-
mistakes in it, or generate a correct chain of reasoning in cases
tation of knowledge and in the emphasis given to the various
too complex to be coped with by humans.
concepts involved.
This article describes the formal basis for reasoning systems,
The language of propositional logic takes the highest level of
points out a number of issues of particular importance for the
abstraction by formalizing knowledge at the level of entire
design of such systems, surveys their numerous applications,
propositions. For instance, the statement “The sun is shining”
addresses the philosophy and psychology of reasoning, and
is taken as an atomic element in the language, thus abstracting
offers some literature on the subject.
from its intrinsic structure altogether. In the formalism the
statement may therefore be abbreviated by any propositional
variable, say S. As seen in section Variants, Specializations and
Abstractions Extensions, it is amazing how far the modeling power of this
relatively poor language can already lead us.
Reasoning takes place in human minds. The understanding of The language of first-order logic (FOL) is less abstract and
the processes in the brain involved in reasoning is still in an hence much more expressive. In our example, it allows the

International Encyclopedia of the Social & Behavioral Sciences, 2nd edition, Volume 5 http://dx.doi.org/10.1016/B978-0-08-097086-8.43036-9 933
934 Deductive Reasoning Systems

distinction between the subject “The sun” in the noun phrase say M, which guides the process of generalizing from given
from the property “is shining” expressed in the verb phrase (or observations. Taking this into account, inductive reasoning
predicate) as well as the modeling of their relationship in actually boils down exactly to deductive reasoning, namely, in
a functional way, e.g., as is_shining(The_sun). In this way, it lays the form K,M,O E whereby now the explanation E is deduced
emphasis on the functional, or - after the philosopher Gottlob from K and O guided by M in some way.
Frege - Fregean, relationships among relations, properties, and In summary, all prevailing modes of human reasoning may
objects. Many knowledge representation formalisms are first- actually be regarded as deductive reasoning. This explains why
order formalisms in disguise. Among those are the Aristotelian deductive reasoning systems have such a great importance for
language, which emphasizes sets rather than functions, the modeling inference in all its generality.
language of description logics (Bader, 2003) emphasizing
concepts, associative nets emphasizing the relationships
Syntactic Inference
among objects and concepts in a graphical representation, and
so forth. So FOL is probably the most popular formal The fundamental observation already made more than two
language for knowledge representation, especially in the thousand years ago by Aristotle and his predecessors is that
context of reasoning systems. reasoning (apart from psychological considerations) is
Modal, transitional, and higher order languages including a matter of form rather than of content. This became apparent
lambda calculus extend the expressiveness of first-order once knowledge was represented in a formalized way. The
languages even further. For the purpose of this article we put knowledge “apes are mammals”, in the sense of “whenever the
the emphasis on FOL, which includes propositional logic as property of being an ape holds then also the mammal prop-
a special case. Most of our considerations generalize in some erty holds”, or A/M, and “mammals have hairs”, M/H,
way to the other more general languages. allows us to conclude that “apes have hairs”, A/H. The
inference is valid independently of the meaning of the words
apes, mammals, and hairs, or of the symbols A, M, and H for
The Inferential Relationship
that matter. Rather it is exclusively the form of the knowledge
Reasoning takes place with a given body of knowledge, say K, that sanctions the validity of such an inference. This funda-
already available. In the predictive mode of reasoning we want to mental property of reasoning enables machines, which to date
determine what facts, say P, could be true in the future. Correct are ignorant of the meaning of words or symbols, to perform
reasoning distinguishes between facts possible on the basis of K correct inferences.
and those that are impossible. In other words, there is a relation The example demonstrates a fundamental rule of inference:
that associates K and P in the positive case. Following Frege this A/M and M/H imply A/H. It is called (a restricted form of)
relation is formally denoted by . That is, K P is short for resolution or (a slightly generalized form of) modus ponens. As we
expressing that P is a correct reasoning prediction on the basis of see, it produces new knowledge (A/H) by eliminating the
K. In the predictive mode of reasoning, we thus want to determine intermediate atom M; thus, the rule may also be seen as
P such that K P holds, or test for a given P whether it holds. implementing a form of transitive closure.
Human reasoning is mostly done with respect to chrono- In order to program computers to search for chains of
logical processes and to a significant extent also concerns inferences we first need rules of inference such as the one just
knowledge about actions and changes taking place in time. On discussed in addition to the formal languages mentioned
the other hand, a large part of the inferential relation is so to before. The combination of a formal language along with a set
speak “timeless”. To simplify further, most logical studies have of rules is called a (formal or logical) calculus. If F is some
abstracted from the issues concerning time and change. Only in formula, i.e., a propositional expression in the formal language
the last few decades have those issues become of great interest under consideration, viz. FOL according to our convention,
within Intellectics, the field studying intelligent behavior and and G is a formula obtained from F by applying the rules of the
systems within Artificial Intelligence (AI), Cognitive Science, calculus, then we write F G. Since is dependent on the
and Neuroscience. The result of those recent studies is that the calculus, say c – which, as in the case of formal languages, can
timeless part of is already expressive enough to model the be chosen from a great variety of alternatives – we should rather
part involving time as well and thus the entire relation . write c but omit the index whenever the calculus is clear from
Insofar this abstraction from time is without restriction of the context. We want to design calculi in such a way that F G
generality. Under this abstraction the predictive mode of holds if, and only if, (shortly, iff) F G, in which case we speak
reasoning is called the deductive mode of reasoning. Because of correct (if) and complete (only if) calculi. From now on we
of what we just said, deductive reasoning is timeless reasoning assume that the rules are chosen such that the resulting calculi
at the surface, but may model general predictive reasoning. are always correct and complete.
In the explanatory mode of reasoning based on K, we start We go one step further in the abstraction process described
with some observation O and basically look for some expla- in this entire section. According to the well-known deduction
nation E so that K,E O. This mode of reasoning is called theorem, it holds that F G iff F/G. By (iteratively)
abductive or inductive depending on the circumstances con- applying this theorem, we can always achieve that there are no
cerning O and E. Since is a relation, the difference to formulas to the left of . Most deductive systems in use are
deductive reasoning is minor; instead of the second argument therefore systems that test whether, for some given calculus
of the relation here just (part of) the first one is to be and some given formula F, it holds that F (establishing
determined. The relationship is even closer. In abductive or an affirmative proof), or dually, that :Fh false signaling
inductive reasoning, we employ additional metaknowledge, a contradiction (proof by contradiction).
Deductive Reasoning Systems 935

Let us illustrate the result of this abstraction process by our clause :H then the empty clause results as resolvent. The empty
apes–mammal–hair example. The example asserts that A/M clause represents false. Its occurrence in this process is the
and M/H implies A/H. Using the logical symbol ^ instead criterion for success. Once this criterion is fulfilled it is guar-
of “and” the assertion reads (A/M) ^ (M/H) (A/H). anteed that the original formula, from which the deduction
Applying the deduction theorem yields (A/M) ^ (M/H) process started off, is indeed valid.
/ (A/H). The formula to the right of is an example for A deductive system based on resolution works as a satura-
what was briefly denoted by F above. It could be presented as tion process. It takes the given clause set, and iteratively applies
input to a deductive system, which would easily verify that the resolution until the empty clause is obtained. The advantage of
formula is valid. If the system works by contradiction, the this type of system is the simplicity of the data structures
presented formula would rather be :[(A/M) ^ (M/H) / involved (clause sets) and of the operations performed at each
(A/H)]. iteration (resolution). The problem is that without additional
The field in which such systems are developed, studied, and constraints the system produces relevant and irrelevant inter-
applied is called automated deduction (AD) or automated mediate results alike, and for hard problems often is drowned
reasoning (AR). It has applications in any field where reasoning in up to billions of generated clauses. An irrelevant interme-
of any kind is relevant, which actually is most of science, or diate result in our earlier example might be that “apes have
most of human endeavors for that matter. Since AD’s first eyes” on the basis of the additional knowledge that “mammals
application was to mathematics, the field is also known as have eyes” when our only interest is in the question about
automated theorem proving (ATP). hairs. Strategies (like the so-called set-of-support strategy) and
reduction rules to some extent help to reduce the amount of
irrelevant clauses generated.
Issues in Automated Deduction An alternative to this saturation-based approach are goal-
oriented tableaux-based deductive systems. Human proofs, i.e.,
There are numerous deductive systems that differ in the formal chains of inferences, may be regarded as treelike structures, or
languages or the set of rules making up the underlying calculus. (closed) tableaux, which in each branch start out from an
Further distinctions are the strategies used for the proof search, obviously valid fact. The basic data structure of tableaux-based
the data structures used for the representation of the formal systems is a (not necessarily closed) tableau. Starting with the
constructs involved, the programming language used for the tableau consisting of the (denial of the) given formula, the
code, and so forth. Some of these issues are discussed in the tableau is extended at each iteration with the goal of closing it.
present section. The data structure (tableau) and the operations (extension of
tableau) are more complicated than those in the saturation-
based systems, which complicates the system development. But
Clausal Form, Resolution, Tableaux, and Connection Method
the approach lends itself to a more goal-oriented fine-tuning
A very popular language is a version of the language of FOL, of the search for the final proof and can therefore avoid some
called clause form. It uses sets of clauses. A clause is a disjunction of the redundancy involved in resolution systems without
of literals. A literal is an atom or its negation. An atom is the extra efforts.
simplest propositional construct such as is_shining(The_sun) or Systems based on the connection method may be regarded as
the A in the ape example, i.e., without involving any logical a compactly coded form of tableaux systems. In this method,
connections. all information coded by a tableau is attached in an extremely
Any formula of FOL can be transformed into clause form by condensed form to the parts of the given formula as structural
a straightforward algorithm based on well-known laws of information. Starting with an initial structure of this kind, only
propositional logic (and techniques discussed in the next the structure is modified at each iteration until a certain final
section); the algorithm is not explained here in detail. Its appli- structure signaling validity of the formula is reached; in contrast
cation to our formula :[(A / M) ^ (M / H) / (A / H)] to the previous kinds of methods, the formula itself is not
leads to the following set of four clauses: {(:A n M), (:M n H), changed at all in the process. As we saw already in the resolu-
A, :H}. The meaning of the comma separating the clauses in tion method, connections (i.e., complementary pairs of
the set may be viewed as that of a conjunction ^. literals) play a fundamental role in proofs. In the connection
A correct and complete calculus based on clause sets is given method they guide the proof search. In the final structure just
by (essentially) the single rule of resolution, which was mentioned, a so-called spanning set of connections for the
mentioned already in section Syntactic inference. In clause formula is identified, which characterizes the validity of the
form, it operates on two selected clauses in the given clause set; formula. For our example, this structure consisting of three
the two clauses must be of the form 1 n A n 2 and 3 n :A n spanning connections is illustrated in the following picture.
4, where the numbers stand for arbitrary, possibly empty
clausal parts. The pair {A, :A} is called complementary, or
a connection. The result of a resolution step then is 1 n 2 n 3 n
4, called the resolvent. For example, the resolvent of the first
two clauses :AnM and :MnH in our example is :AnH. The
resolvent (of each resolution step in the derivation) is added to
the set of clauses. If the next resolution step selects this last Research in AD pursues all three kinds of approaches, with
resolvent and the clause A, then the newly added resolvent the focus in each case being on eliminating redundancy. This
thereof is H. If this resolvent again is selected together with the can be achieved in various ways. One way consists of refined
936 Deductive Reasoning Systems

calculi in terms of the rule sets used. For instance, the appli- described. For instance, the resolvent of the first two clauses in
cation of resolution may be restricted in many ways without the apes example now is :AxnHx and similarly for the two
sacrificing completeness (e.g., linear resolution). This also other steps. Except for the additional unification the proof
includes the application of reduction rules (e.g., subsumption). proceeds as before, resulting in the empty clause. In general, the
Another way to reduce redundancy is by applying various resolution method requires an additional rule of inference, viz.
strategies that guide the rules application. A third is by per- factoring, by which two different literals in a clause by way of
forming several small steps as a single global step (e.g., unification are factored into a single one.
hyperresolution), which again narrows the search space. A The proof of our running example in the connection
fourth involves global steps sanctioned by theories involved method is unchanged as well:
(such as equality theory, see section Variants, specializations
and extensions). Finally, ever improved software engineering
techniques and faster machines contribute to a continuous
enhancement of system performance. On top of all, heuristics
try to achieve success faster by sacrificing completeness of the
method. Note that each connection corresponds to a respective step
in the resolution derivation described above. For comparison,
First-Order Features recall that resolution in each step adds a new clause to the
clause set and thus with it new connections. In contrast, the
For simplicity, the introduction of some AD methods in the connection method is concerned exclusively with the connec-
previous subsection has glossed over a number of details. tions present in the formula to be proved, although these
While speaking of FOL, no first-order features beyond the connections by the nature of FOL may be required in different
propositional ones were actually taken into account. For instances. This allows a more focused search, guided by the
instance, the apes example actually talks about objects with structure of the formula in a connection-oriented way. In
certain properties so that its first-order formula would rather addition, in the connection method no transformation into
read cx [(Ax/Mx) ^ (Mx/Hx) / (Ax/Hx)]. In order to clause form is required (Otten, 2011). This is an advantage
save parentheses, the atoms like Ax are short for the functional since this transformation in general introduces redundancy
form A(x) reading “x is an ape”. The statement is asserted for all thus increasing the search space. On the contrary, the given
objects x, formally expressed by the universal quantifier cx. In formula may even be reduced by factoring, making it more
addition, FOL features an existential quantifier as in the sen- nonnormal, possibly resulting in (much) shorter proofs and
tence “any human has a mother”, formally cxdy (Hx/Myx), less proof search (Bibel, 2006).
whereby Hx is short for “x is human” and Myx abbreviates “y is In section Syntactic inference we have made the distinction
mother of x”. between affirmative proofs and proofs by contradiction. Tech-
As these examples illustrate, atoms in FOL have an internal nically, there is no argument for one or the other form. Any
structure. It consists of a predicate symbol like M and a certain method can be formalized either way without affecting
number of arguments, two in the case of M. The arguments can performance. In addition, there is an “affirmative” clause form
be variables like x or constants like “The_sun”, or shortly s. They defined similarly to the one introduced in section Clausal form,
may even involve functions like “the color of s”, or shortly c(s). resolution, tableaux, and connection method. Resolution by
Altogether we speak of terms serving as arguments. historical accident is usually presented as a proof-by-contra-
The occurrence of quantifiers in the formulas to be deduced diction method, while the connection method is usually
affects the transformation algorithm into clause form. By affirmative. The tableaux method is used both ways.
a technique called skolemization existential quantifiers are
eliminated and then all-quantifiers ignored by default. As
Variants, Specializations, and Extensions
a result, a set of clauses looks as before except for the more
complicated atoms involved. For instance, the apes example FOL is a rather expressive formalism in which, for instance, all
would now look like {(:AxnMx), (:MxnHx), Ax, :Hx}. of mathematics as well as more or less all of science can be
Connections like {A, :A}, because of the richer structure of represented. Typically, deductive tasks arise within a given
atoms, may now become more complicated such as e.g., theory. For instance, if the task F involves an equation then the
{A(x,f(z)), :A(b,y)}. Whether or not such a connection can be theory has to contain the equality axioms, say T¼ . These are, for
made complementary and thus will contribute to a proof is subject instance, the axioms of reflexivity and substitutivity (for func-
to a further technique called unification, which assigns values (or tions and predicates). The formula to be proved thus has the
rather terms) to the variables involved. In the present example, x form T¼/F, and it is straightforward for deductive systems to
would be assigned the term b, and similarly y the term f(z); in attempt a proof for such a formula.
short, this substitution may be written as {x\b, y\f(z)}, resulting in In comparison with the human way to solve equations,
the complementary connection {A(b,f(z)), :A(b,f(z))}. however, this approach involving the equality axioms seems
The three types of methods introduced in the previous pretty awkward. Rather, humans proceed in such cases by
subsection engage unification in some way, which will not be substituting terms by equal terms until the equality becomes an
described here in detail. Resolution, enhanced by unification, identity. This is the motivation for treating equality in rather
allows the elimination of intermediate literals (i.e., connec- special ways also in deductive systems. For instance, resolution
tions) even if they are not exactly identical but can be made has a special rule called paramodulation by which equal terms
identical by replacing variables in the atoms by terms as just may be substituted without any reference to axioms. Similar
Deductive Reasoning Systems 937

techniques have been developed for the tableaux and the over functions and predicates according to familiar practice,
connection method. In fact, there is an extensive research area e.g., in mathematics. A formalism for representing HOL is type
known as “term rewriting”, in which efficient algorithms for theory. For all these extensions there are deductive formalisms
treating equality theories have been developed. Appropriate and systems, analogous to those of FOL, yet more complicated.
techniques are also available for other special mathematical All formalisms discussed above test the validity of formulas,
tools (like mathematical induction) and theories. i.e., their truth under any interpretation. Often it is already
The programming language PROLOG (Clocksin and Mellish, sufficient to test the given formula for its satisfiability, i.e., its
2003) is a restricted FOL language enriched by control features. truth under one particular interpretation rather than under all
It restricts the formulas to so-called Horn clauses, which in interpretations. This semantic approach has found many
the contradiction mode means that there is at most one negative applications especially for the restricted language of
literal in each clause. A PROLOG clause is written propositional logic. The satisfiability of propositional formulas
A0 :¼ A1,A2,.,Am, with the left side of :¼ called the head and is known as the problem class SAT. For instance, the apes–
the right the body of the clause. For instance, the clause Ax/Mx mammal–hair example formula (A/M) ^ (M/H) /
from the apes example here reads Mx :¼ Ax, which by (A/H) will turn out true if all propositional variables A, M, H
comparison of the two versions shows that :¼ can be interpreted as are assigned the value true (and in fact under any assignment
an implication from right to left. In the case m ¼ 0 (i.e., no body) due to the formula’s validity).
the sign :¼ is deleted. A PROLOG program is a set of PROLOG There are extremely powerful systems for SAT problems and
clauses. A clause with no negative literal is written ?– A1,A2,.,Am. variants of these, solving problems that may involve millions of
It is called a query. Except for these superficial syntactic differences propositional variables. One of the popular techniques uses
a PROLOG system behaves much like a deductive system based binary decision diagrams, or BDDs (Bryant, 1992). This formalism
on the methods discussed before. One exception is the possibility treats propositional formulas, restricted to the operators
of allowing negated atoms not Amþ1 in the clause body, which conjunction, disjunction, and negation, in a functional way.
are treated with a special technique called “negation as failure”. Here conjunction is written like a product, disjunction as sum, and
In addition, with respect to applications in AI the semantics of negation with a bar, e.g., f(A,M,H) ¼ [AM þMHþ 
 AþH]þ  þ H]
A
a PROLOG program differs from that of a corresponding classical for the formula after elimination of / according to the rules of
formula. The control features allow the programmer to overrule propositional logic. The efficiency of BDDs is achieved basically
the system’s built-in control. by a reduction in the formula by factoring identical subformulas
In section Formal languages we already mentioned and by choosing a special order of the variables in the formula.
description logic (DL) as a first-order formalism in disguise. Answer set programming (Gelfond, 2008), or ASP, is a form
More precisely, DL is actually a family of formal knowledge of declarative programming similar to PROLOG, but with
representation languages which, due to certain restrictions, a different semantics and solution technique. The language
feature more efficient deductive proof procedures than full FOL uses PROLOG clauses with a number of additional features.
(mostly even decision procedures). It models concepts, roles The semantics is based on so-called stable models. The
and individuals, and their relationships. For instance, woman ¼ solution technique includes a grounder and a solver. The
adult gender:female might be a definitional statement of the grounder in essence transforms the problem into
concept “woman” in DL. In FOL such a concept is represented a propositional formula; the solver is something like a SAT
as a predicate (like in Woman(mary)). This syntactic difference solver. Among the additional features are several language
is irrelevant from an efficiency point of view; the decisive extensions like integrity constraints, choice rules, cardinality
difference lies in the restrictions imposed on the operations of and weight constraints, and so forth. These language features
the language, which lead to less expressiveness of the language render ASP a rather expressive representational formalism.
but to a higher efficiency of the resulting deductive systems. ASP is extremely powerful for solving difficult search
Another restricted form of FOL is the formalism of syllo- problems. Due to its declarative nature, it offers a great
gisms, originally invented by Aristotle. It features judgments of advantage in comparison to programming such problems in
the form all apes are mammals, or AaM, at least one ape is old, or a conventional programming language.
AiO, and so forth. Along with judgments of these kinds there
are inference rules. For instance, the apes–mammals–hair
Complexity, Systems, and Performance
example is an instance of what is known as the valid inference
Barbara: AaM, MaH / AaH. The traditional names of these From a theoretical point of view, AD is one of the hard prob-
rules are chosen for mnemotechnical reasons: the three a’s in lems in computer science. SAT solving, i.e., semantic proposi-
the word “Barbara” remind of the three a’s in the rule. In tional reasoning, is already NP-complete, which (for the state
Brüning (1996) a strict formalism based on syllogisms has of art) essentially means no procedure can solve all proposi-
been developed but not yet been realized in a running system. tional formulas efficiently, even if the formulas are comparably
So far we have discussed restricted forms of FOL. small. First-order reasoning is even more difficult, since FOL is
Conversely, there are numerous extensions of FOL, too many undecidable: there is no procedure that can decide for all
to discuss here in any detail. We just mention the great variety formulas if they are valid or not.
of modal and temporal logics that feature additional logical Nevertheless, research in AD has produced a variety of
operators (e.g., the necessity operator , or the possibility reasoning techniques that can solve most reasoning problems
operator >) beyond those of FOL along with corresponding very efficiently. Propositional reasoning techniques like conflict
rules of inference. Higher order logic (HOL) is the most analysis, clause learning, backjumping, adaptive branching,
powerful extension beyond FOL. It allows the quantification random restarts, and problem reductions have increased the
938 Deductive Reasoning Systems

performance of SAT solvers by many orders of magnitude, Mathematical Theorem Proving


which makes it possible to apply them to formulas with
Originally, mathematics has been the primary field of appli-
millions of clauses and variables. In first-order reasoning,
cation for AD. Deductive systems have been applied to
indexing techniques, efficient data structures, and search
virtually all subdisciplines in mathematics, and tens of
heuristics have led to efficient proof systems capable of dealing
thousands of theorems have been proved automatically.
with complex reasoning tasks.
Automated theorem provers have found new proofs for
In the past years, benchmark collections have made it
known theorems as well as proofs of unsolved mathematical
possible to evaluate the performance of AD systems, while
problems and conjectures that had never been proved by
annual system competitions in SAT solving (The International
humans before.
SAT Competitions Web page) and first-order theorem proving
For instance, in 1996, a system called Otter/EQP solved the
(The CADE ATP System) have resulted in significant improve-
so-called Robbins problem, which had been open for more
ments. Currently, the best SAT solvers are 3S, ppfolio, and the
than 60 years. The machine-generated proof consists of only 13
ASP system clasp, while systems like Vampire, Waldmeister,
equations and can therefore be checked easily by human
Spass, E, and leanCoP are the most successful first-order theorem
mathematicians. Finding this proof, however, took more than
provers. However, the winners of these competitions keep
7 days of intense machine computation, as several thousand
changing, since the development of reasoning techniques
equations had to be generated and combined in order to
makes previously unknown systems even more efficient than the
construct the proof (Kolata, 1996, McCune, 1997).
winners of the past years.
Another famous example is the proof of the four color
theorem, which has been conjectured in 1852 and was the
Artificial Neural Systems first theorem to be proved by a computer in 1976 (Appel and
Haken, 1977). To prove the theorem, a special-purpose
Human reasoning takes place in brains. The mechanisms
computer program had to analyze 1936 different maps.
occurring in brains are modeled by connectionist systems. So
This resulted in a proof that is infeasible for a human to
the question naturally arises whether reasoning could be
check by hand and is not accepted as a convincing proof
modeled by a connectionist, also termed artificial neural,
by some mathematicians. In 2005, the theorem could be
system. Since we do not yet know how the brain realizes
proved with the help of a general purpose theorem prover
reasoning, such a system could be a very first step toward
Coq (Gonthier, 2008). Although even this proof cannot be
understanding the brain’s processes in reasoning.
checked without the help of a computer, it was sufficient to
Several neural systems have been developed that perform
dispel remaining doubt about the validity of the four color
deductive (and abductive and other kinds of) inferences,
theorem.
mostly for propositional logic. The so-called CORE method
(“Connectionist mOdel generation using Recurrent networks
with feedforward core”) achieves such inferences semantically
Hardware and Software Verification
by generating or approximating a (possibly least) model of the
premises under which a conclusion can be drawn. In the One of the most important applications of AD techniques
propositional case, the model can iteratively be computed by today is the verification of software and hardware. In the early
a three-layer feed-forward core (i.e., connectionist network) 1990s, major flaws in hardware design such as the Pentium
until a stable state is reached. Thereby, the output nodes of the FDIV bug resulted in hundreds of millions of dollars of
network are recurrently connected back to the input nodes to damages and caused the industry to look for tools that could
allow for the iterations. With a number of additional ideas this validate the design of complex hardware circuits. At the same
method has been generalized to the first-order case (Baader time, new techniques for eliminating redundancy in proposi-
et al., 2008). Here, the model can only be approximated, tional reasoning led to significant efficiency improvements,
however, with arbitrary precision. which permitted the use of model checkers and SAT solvers in
The CORE method bridges the gap between the discrete large-scale applications involving millions of components
world of logic programs and the continuous world of neural (Clarke et al., 1999). As a result, model checkers were adopted
networks. It integrates distinct functions such as transforming as the standard technique to verify the design of computer
programs (as background knowledge) into networks, learning chips before they go into production.
and optimizing networks, generating or approximating models While the verification of hardware involves the analysis of
from networks, drawing inferences by way of the models, a huge number of possible states, the problem is finite and can
and extracting a logic program from a learned or optimized therefore be represented in propositional logic. In contrast,
network. software can assume infinitely many different states, which
makes software verification much more complex. For specific
types of software such as compilers, state-based systems like
Applications communication and security protocols, and some commercial
applications, it has been demonstrated that automated proof
As pointed out in the introduction, reasoning is a fundamental systems and software model checkers are capable of con-
human capability. Therefore, it pervades many of our activities. structing proofs that guarantee the software’s validity. But in
In consequence, reasoning systems have numerous and widely general these systems do not scale well enough to deal with
varied applications. Some of these are briefly discussed in the the complex formulas that are needed for the description of
present section. real-world software.
Deductive Reasoning Systems 939

Software and Hardware Synthesis coping with the masses of information will only be possible if
inductive systems will extract knowledge from the data.
An alternative approach to software validity is based on the
idea that programs and proofs of their correctness should be
developed simultaneously instead of verifying software after it
Philosophy and Psychology of Reasoning
has been created. Program synthesis aims at using AD to
develop “correct-by-design” software from descriptive specifi-
As seen in the previous section, deductive systems are
cations. Ideally, a programmer would focus on designing
becoming a part of daily life. We want them to behave ratio-
a precise formal specification of a given problem and then call
nally in the human sense. Hence, we need to understand
a deductive system to generate the actual implementation. The
human rationality and reasoning in the first place, a goal
programming language PROLOG basically works as such
pursued for more than two thousand years first in philosophy,
a deductive system. However, as a programming language, it
then in psychology, and more recently in intellectics. In this
also contains nondeclarative control mechanisms and is only
section, we discuss philosophical and psychological aspects of
a first step toward “programming in logic”.
reasoning.
Systems that synthesize programs from descriptive specifi-
cations involve strong deductive components to ensure the
correctness of the generated program. The systems KIDS and Can Reasoning Be Mechanized?
Specware/Isabelle have been used in the automated construc-
tion of commercial software and scheduling algorithms that are Logic and the study of deductive reasoning was a primary topic
not only correct by design but also significantly more efficient in philosophy for more than two thousand years (for an
than hand-crafted versions (Smith, 1991, Smith and Parra, exposition of this history see Kneale and Kneale, 1984). Since
1993). The higher order theorem prover Nuprl has been used the nineteenth century, it transformed more and more into
for the semiautomated development and optimization of a mathematical subject, whereas in the past 60 years it has
distributed software with more than 50 000 lines of code (Liu become a major topic in AI. Therefore, today the major
et al., 1999). A key to the success of these systems was the contributions to the subject do not originate in philosophy
fact that the deductive components were guided by strategies anymore – with one exception: many philosophers have tried
that incorporate knowledge about algorithmic structures and in thousands of papers to argue in favor of the thesis that the
efficiency optimizations. mechanization of reasoning has strict limitations and will
In hardware design, the standard procedure consists of never be able to reach the human level of performance.
describing the behavior of circuits in a hardware description Herbert Dreyfub, John R. Lucas, Roger Penrose, and John
language at logic level followed by extensive verification and Rogers Searle are some of the philosophical spokesmen in
logic synthesis, a process of turning the circuit descriptions into this debate. The debate is a fruitless one, and the arguments
a design implementation in terms of logic gates. This procedure in favor of the thesis have proved faulty. One of the papers
consumes significant amounts of time and needs a lot of effort. that have defeated the thesis is LaForte (1998). Although we
While the idea of synthesizing hardware circuits from logical indeed do not know in detail how the brain realizes our
specifications has been proposed several decades ago, it was thinking, the computationalism underlying intellectics at
deemed impractical because of its high computational present is the best available scientific approach.
complexity. Recent developments (Bloem, 2007), however, have
made it applicable to realistic examples, and it is expected that
Psychology of Reasoning
hardware synthesis based on AD will eventually be used in
industrial applications. In many respects, the best available deductive reasoning system
is still the human mind. Psychological studies and thus our
understanding of (the phenomenological part of) human
Knowledge Systems reasoning can therefore enhance the design of artificial
deductive systems (Evans, 1970). The key idea is that humans –
Reasoning generates new knowledge from given knowledge.
in contrast to machines – construct internal representations of
Hence, it is an essential feature in any system that incorporates
the state of affairs described by a logical argument in order to
explicitly represented knowledge. Such a system is variably
understand it. These “mental models” are similar to the
called knowledge system, knowledge-based system, expert
models that people use to perceive the world, but
system, semantic web, ontologies, and so forth. In the CYC
rudimentary and incomplete. Experimental evidence seems to
system an entire encyclopedia of knowledge is encoded in this
indicate that human deduction depends less on laws of
way. Numerous special systems encode knowledge of specific
inference but rather on a search for alternative models of the
disciplines, allow the manipulation of such knowledge in
premises that could refute a given conclusion and that
a deductive way, or can answer questions in complex scenarios
a conclusion is accepted as long as there is no refutation for it.
(Harmelen et al., 2007). For instance, in 2011, the system
Most humans, for instance, appear to read an implication as
WATSON outperformed the two best human players in the US
bi-implication, or equivalence, and draw conclusions like the
quiz show Jeopardy. Similarly, the Apple iPhone 4S features
following.
a system Siri, which understands and answers queries in spoken
natural language, not least because of deductive mechanisms. l If John is in New York then Cathy is in New York too.
These are just three examples out of tens of thousands of such l John is not in New York.
systems. In the emerging technology of real-world awareness, l Therefore Cathy isn’t in New York either.
940 Deductive Reasoning Systems

which obviously is not a correct deduction from the point of Bibliography


view of classical logic like FOL.
Negation is even more difficult. The sentence Anderson, J.R., 1993. Rules of Mind. Erlbaum, Hillsdale, NJ.
Appel, K., Haken, W., 1977. Solution of the four color map problem. Scientific American
l John wouldn’t have been in New York if Cathy hadn’t been there. 237 (4), 108–121.
Baader, F., Hitzler, P., Hölldobler, S., 2008. Connectionist model generation: a first-order
is logically equivalent to the first one above but is perceived approach. Neurocomputing 71, 2420–2432.
differently. Bader, S., et al., 2003. The Description Logic Handbook: Theory, Implementation and
The theory of mental models aims at capturing the differ- Applications. Cambridge University Press, Cambridge, UK.
ence between human and formal reasoning. In this theory, the Bibel, W., 1993. Deduction: Automated Logic. Academic Press, London.
Bibel, W., 2006. Research perspectives for logic and deduction. In: Reasoning, Action,
process of deductions depends on a semantic procedure that
and Interaction in AI Theories and Systems, LNAI, vol. 4155. Springer, pp. 25–43.
constructs an internal representation of what has been stated, Bibel, W., Schmitt, P.H., 1998. Automated Deduction – A Basis for Applications. In:
tries to draw possible conclusions by formulating a more Applied Logic Series, vols. I-III. Kluwer, Dordrecht, NL.
elaborate model based on that representation, and attempts to Bloem, R., et al., 2007. Automatic hardware synthesis from specifications: a case study.
find a countermodel for that conclusion. If no such counter- In: Conference on Design, Automation and Test in Europe, pp. 1188–1193.
Brüning, W., 1996. Grundlagen der Strengen Logik. Königshausen und Neumann,
model is found, the conclusion is considered valid and the Wurzburg.
process continues by proposing further conclusions and trying Bryant, R., 1992. Symbolic boolean manipulation with ordered binary-decision diagrams.
to refute them. ACM Computing Surveys 24 (3), 293–318.
On this basis, cognitive scientists and psychologists have Clarke, E.M., Grumberg, O., Peled, D., 1999. Model Checking. MIT Press,
Cambridge, MA.
formulated a comprehensive theory of human inference that
Clocksin, W., Mellish, C., 2003. Programming in Prolog, fifth ed. Springer, Heidelberg.
explains propositional, quantificational, and relational reasoning Evans, J.St.B.T., 1970. The Psychology of Deductive Reasoning. Routledge & Kegan
(Johnson-Laird and Byrne, 1991, Johnson-Laird, 1999). This Paul, Reading, UK.
theory has been used as a foundation for implementing reason- Fitting, M., 1990. First-Order Logic and Automated Theorem Proving. Springer,
ers that appear to be a more appropriate model of the human Heidelberg.
Gabbay, D., Hogger, C., Robinson, J.A. (Eds.), 1991. Handbook of Logic in Artificial
intellect. There are also theories that view human deduction as Intelligence and Logic Programming. Oxford University Press, Oxford, UK.
a process based on factual knowledge (Anderson, 1993, Newell, Gelfond, M., 2008. Answer sets. In: Handbook of Knowledge Representation. Elsevier,
1990) or on formal rules of inference (Rips, 1994) similar to Amsterdam, pp. 285–316.
those used in AD. Evidence from a wide range of studies Gonthier, G., 2008. Formal proof – the four-color theorem. Notices of the American
Mathematical Society 55 (11), 1382–1393.
seems to indicate that the theory of mental models provides the
van Harmelen, F., Lifschitz, V., Porter, B. (Eds.), 2007. Handbook of Knowledge
most appropriate representation of the mechanisms underlying Representation. Elsevier, Amsterdam.
human deduction. However, in contrast to other areas in psy- Johnson-Laird, P.N., 1999. Deductive reasoning. Annual Reviews Psychology 50,
chology such as image processing (Kosslyn, 1995) or “system 1 109–135.
mental operations” (Kahneman, 2011), reasoning (a system 2 Johnson-Laird, P.N., Byrne, R.M.J., 1991. Deduction. Lawrence Earlbaum Associates
Ltd, Hillsdale, NJ.
operation) is still poorly understood. Kahneman, D., 2011. Thinking, Fast and Slow. Farrar, Straus and Giroux, New York.
Fitting, 1990 is an excellent textbook on propositional and Kneale, W., Kneale, M., 1984. The Development of Logic. Clarendon Press, Oxford, UK.
first-order logic. Bibel, 1993 is a special textbook on all Kolata, G., December 10, 1996. Computer Math Proof Shows Reasoning Power. The
aspects of deductive methods and systems. Bibel and Schmitt, New York Times. http://www.nytimes.com/library/cyber/week/1210math.html.
Kosslyn, S.M., 1995. Image and Brain – The Resolution of the Imagery Debate. MIT Press,
1998 is a reader covering all modern deductive techniques,
Cambridge, MA, pp. 327–404.
a number of very successful systems, and several important LaForte, G., Hayes, P.J., Ford, K.M., 1998. Why Gödel’s theorem cannot refute com-
applications. Gabbay et al., 1991 is a handbook that includes putationalism. Artificial Intelligence 104, 265–286.
deductive reasoning systems and other topics. Robinson and Liu, X., et al., 1999. Building reliable, high-performance communication systems from
Voronkov, 2001 is a handbook covering just deductive components. Operating Systems Review 33, 80–92.
McCune, W., 1997. Solution of the Robbins problem. Journal of Automated Reasoning
methods, techniques, and systems. Plaisted, 2009 is an 19 (3), 263–276.
encyclopedia article similar to the present one. Ongoing Newell, A., 1990. Unified Theories of Cognition. Harvard University Press, Cambridge, MA.
research is reported in the Journal of Automated Reasoning Otten, J., 2011. A non-clausal connection calculus. In: TABLEAUX 2011, LNAI 6793.
(Kluwer Academic Publishers) and in the proceedings Springer, pp. 226–241.
Plaisted, D., 2009. Theorem proving. In: Encyclopedia of Computer Science and Engi-
volumes (published by Springer Verlag) of the International
neering. John Wiley and Sons, New York, pp. 223–244.
Conferences on Automated Deduction as well as in other Rips, L.J., 1994. The Psychology of Proof. MIT Press, Cambridge, MA.
journals and conferences. Robinson, J.A., Voronkov, A. (Eds.), 2001. Handbook of Automated Reasoning, vols. 2.
Elsevier Science and MIT Press, Amsterdam, NL.
Smith, D.R., 1991. KIDS – a knowledge-based software development system. In:
See also: Artificial Intelligence: Connectionist and Symbolic Automating Software Design. AAAI Press/The MIT Press, pp. 483–514.
Approaches; Knowledge Representation; Nonstandard Smith, D.R., Parra, E.A., 1993. Transformational approach to transportation scheduling.
8th Knowledge-Based Software Engineering Conference, 60–68.
Reasoning; Organizational Decision Making; Problem Solving The CADE ATP System Competition. (http://www.cs.miami.edu/~tptp/CASC)
and Reasoning: Case-based. The international SAT Competitions web page. (http://www.satcompetition.org)

You might also like