Int. J.

Man-Machine Studies (1986) 24, 65-75

Fuzzy cognitive maps

BART KOSKO
VERAC, Incorporated, 9605 Scranton Road, Suite 500, San Diego, CA 92121, U.S.A.

(Received 15 April 1985, and in revised form 10 September 1985)

Fuzzy cognitive maps (FCMs) are fuzzy-graph structures for representing causal reason-
ing. Their fuzziness allows hazy degrees of causality between hazy causal objects
(concepts). Their graph structure allows systematic causal propagation, in particular
forward and backward chaining, and it allows knowledge bases to be grown by
connecting different FCMs. FCMs are especially applicable to soft knowledge domains
and several example FCMs are given. Causality is represented as a fuzzy relation on
causal concepts. A fuzzy causal algebra for governing causal propagation on FCMs is
developed. FCM matrix representation and matrix operations are presented in the
Appendix.

1. Introduction: the knowledge acquisition/processing tradeoff

Most knowledge is specification of classifications and causes. In general, the classes
and causes are uncertain (fuzzy or random), usually fuzzy. This fuzziness passes into
knowledge representations and on into knowledge bases, where it leads to a knowledge
acquisition/processing tradeoff. The fuzzier the knowledge representation, the easier
the knowledge acquisition and the greater the knowledge-source concurrence. But the
fuzzier the knowledge, the harder the (symbolic) knowledge processing.
Fuzzy cognitive maps (FCMs) circumvent the tradeoff. FCMs are fuzzy-graph
structures for representing causal reasoning. Their fuzziness allows hazy degrees of
causality between hazy causal objects (concepts). Their graph structure allows system-
atic causal propagation, in particular forward and backward chaining, and it allows
knowledge bases to be grown by connecting different FCMs. FCMs are especially
applicable in soft knowledge domains (e.g. political science, military science, history,
international relations, organization theory) where both the system concepts/relation-
ships and the meta-system language are fundamentally fuzzy.

2. Cognitive maps
Political scientist Robert Axelrod (1976) introduced cognitive maps in the 1970s for
representing social scientific knowledge. Axelrod's cognitive maps are signed digraphs.
Nodes are variable concepts (like social instability, not like society) and edges are
causal connections. A positive edge from node A to node B means A causally increases
B. A negative edge from A to B means A causally decreases B. Cognitive maps facilitate
documentary coding, constructing symbolic representations of expert documents (Henry
Kissinger's documents code well; see Fig. 1).
Axelrod exploited the (adjacency) matrix representation of cognitive maps (see Fig.
2). Causal conceptual centrality in cognitive maps can be defined with adjacency-matrix
65
0020-7373/86/010065 + 11503.00/0 t~) 1986 Academic Press Inc. (London) Limited
 66 a. KOSKO

C~

mentahsm

Im~~ ( rad~rc~lbs' m

SoY;' ~ ba~ rnOI _ ter~oLr

los%

FIG. 1. A cognitive map constructed from Henry A. Kissinger's article "'Starting Out in the Direction of
Middle East Peace" (printed in the Los Angeles Times, Summer 1982). Positive edges represent causal
increase. Negative edges represent causal decrease. The "policy variable" is ISLAMIC FUNDAMENTAL-
ISM. The "value variable" is STRENGTH OF LEBANESE GOVERNMENT. Other concept nodes are
"cognitive variables."

C, Cz Cs C4 C5 C6

Cl 0 -I I 0 o o\
Cz 0 0 0 l 0 0

q o o o o I 0

,//
c4 o o o o 0 -I

c5 o o o -~ o

cs o o o o 0 0

FIG. 2. The adjacency-matrix representation of Kissinger's cognitive map in Fig. 1. eij= e(C~, Ci) is the
causal edge function value, the causality causal concept node C, imparts to Cj. Ci causally increases Cj if
% = 1, causally decreases Cj if % = -1, and imparts no causality if % = 0.
FUZZY COGNITIVE MAPS 67

components and much causal chaining information can be obtained from reachability
matrices. These techniques are reviewed and fuzzified in the Appendix.
In general, cognitive maps are too binding for knowledge-base building. For, in
general, causality is fuzzy. Causality admits of degrees, and vague degrees at that. It
occurs partially, sometimes, very little, usually, more or less, etc. More generally still,
the knowledge-base building promise of cognitive maps is combining knowledge
sources' cognitive maps, but the fuzziness of the combined knowledge rises to the level
of fuzziness of the fuzziest knowledge source. Fuzzy cognitive maps accomodate this
knowledge-base building property. An example is the FCM in Fig. 3, where the causal

Early ~ (Late
mission phase mission phase
action ~ 9 action

Febo
movement ) T Interd'cf'~
act,on

,.Jr

Bridge Verymuch Bridge

use ~ damage

Target
value

FIG. 3. Bridge Target Value FCM. Strategic objectives, mission tactics, and facts of the battlefield are related
by fuzzy causality to produce a net utility effect. (FEBA stands for Forward Edge of Battle Area and hence
as it moves, the probability that a random bridge will be used increases.)

relationships affecting a bridge's tactical target value are fuzzy in the manner of military
science.
The next three sections formally develop FCMs and a fuzzy causal algebra for
propagating causality on a FCM.

3. Representing causal reasoning

David Hume notwithstanding, causality is more complicated than logical implication.
Consider causal increase, or positive causality. If "A causes B" is represented as "A
implies B," then, by contraposition, "A causes B" is everywhere replaceable with
"not-B causes not-A." But though smoking causes lung cancer, not having lung cancer
does not cause non-smoking. Rather what can be inferred is that not smoking tends
to cause lung non-cancer, so to speak. And this is a quite general relationship among
68 B. KOSKO

causally increasing quantities (positive correlates). If A causes B, then increasing A

increases B, and decreasing A decreases B. The inverse relationship holds between
causally decreasing quantities (negative correlates). If A causally decreases B, then
increasing A decreases B, and decreasing A increases B.
The non-implication nature of causal reasoning can be represented in a fuzzy set
(logic) framework. Since causal objects are variable concepts, they can be represented
as fuzzy subsets of some concept space, where change in fuzzy-set membership degree
represents concept variation. Define a concept C~ as the fuzzy union (disjunction) of
some fuzzy quantity set Q~ and associated dis-quantity set ~Q~:

C,=Qiu-Qi.

-Q~ can be thought of as the abstract negation, or local fuzzy set complement, of
0~. The "negation" operator - simply indicates a set partition. (The negation of the
concept is still its set complement.) It need only be a set index for the pair (Qi, -Q~)
that obeys double negation: -~Q~ =Q~. Otherwise Q~ and ~Q~ are arbitrary fuzzy
sets. Fuzzy causality can then be defined in terms of fuzzy set-theoretical (logical)
relationships among fuzzy concepts9 Let Cj = Qj w - Q j .
Definition. Ci causes Cj iff Qi C Qj and - Q i c - Q i ;
Ci causally decreases Cj iff Qi c - Q j and - Q i c Qj, where " c " stands for fuzzy set
inclusion (logical implication).
Hence negative causality can be defined with the same fuzzy quantities and relation-
ships as positive causality; i.e. negative causality is eliminable. Hence the negative
causal relationship

is equivalent to the positive causal relationship

This observation leads to a general rule of replacement in FCM construction.

Rule. Replace every C~ -'3_Cjwith Ci ~-Cj.

Henceforth the negative causal arrow -~ will not occur. The unsigned arrow --, will
mean positive causality. The cognitive maps in Figs 1 and 3 are displayed transfor-
med in Figs 4 and 5.
A second, more subtle point when representing causal reasoning is that modifiers of
causal quantities need not be negated (complemented)9 Not smoking causes lung
dis-cancer, not non-lung cancer. To develop a fuzzy sociological example, observe that
FUZZY COGNITIVE MAPS 69

the cognitive-map fragment

is equivalent to (or, more generally, functionally related to) the fragment

Here the generic causal quantities are T H R E A T and STABILITY (between which
causal decrease still holds). The modifiers are NUCLEAR_WAR and SOCIAL. No
assumption is made that modifier (fuzzy) sets are closed under abstract negation
(complementation) in the underlying cognitive space. In particular, there need be no
DIS_NUCLEAR_WAR or NOT_SOCIAL fuzzy subsets. Hence, more generally, causal
concepts are built out of fuzzy set relations among quantity, dis-quantity, and modifier
fuzzy subsets. Figure 6 pictures this situation in the sociological example.

~x.x.

FIG. 4. Positive-causalityrepresentation of Kissinger's cognitive map in Fig. I. All edge arrows indicate
positive causality (causal increase). Edge arrows from new dis-concepts are dashed.
 70 B. KOSKO

{ mission phose ) [ m,ss,onl3hose ) [ missionphase ) [ m,ss,onPhose )

FIG. 5. Positive-causality representation of bridge target value FCM in Fig. 2. Fuzzy causal weights have
been omitted for convenience. Edge arrows from new dis-concepts are dashed. (The default assumption is
that the fuzzy causal weight between transformed dis-concepts is the same weight as between the untrans-
formed concepts. In general, the weights may differ.)

W
FIG. 6. Fuzzy-set representation of causal concepts. A modifier set fuzzily intersects the fuzzy union of a
quantity and dis-quantity set. (In general, causal concept components need not be connected.) Causal
increase or decrease is represented by the appropriate inclusion of intersections.
FUZZY COGNITIVE MAPS 71

The cognitive-map fragment in the sociological example can be symbolically rep-

resented as set inclusions of set intersections:

T H R E A T n NUCLEAR_WAR c INSTABILITY n SOCIAL,

D I S _ T H R E A T ca NUCLEAR_WAR c STABILITY ca SOCIAL,

where "ca" stands for fuzzy set intersection (conjunction).

More generally, let Q~, ~Q~ and M~ be the respective ith quantity, dis-quantity, and
modifier fuzzy sets. (The default M~ is the set universe.) Then the ith causal concept
C~ can be defined as C~ = (Q~ u ~Q~) n M~. This leads to the final definitions of causal
increase and decrease.

Definition. C, causes Cj iff (Q, n M,) c (Qj n Mj) and ( ~ Q , n Mi) c (~Qj n Mi); Ci
causally decreases Cj iff (Qj n Mi) c (~Qj n Mj) and (~Qi n Mi) c (Qj ca Mj).

4. The abstract FCM framework

FCMs are fuzzy causal graphs (fuzzy graphs). The apparatus needed to define them
requires embedding the concepts and definitions of the previous section in abstract
fuzzy spaces. The fuzzy logical analogue is immediate but will not be mentioned.
Let X be some non-empty set. Let F(2 x) denote the fuzzy power set of X - - t h e set
of all fuzzy subsets of X. For fuzzy subsets A, B e F(2X), define the degree ofsubsethood
(Bandler & Kohout, 1980) of A in B by me(2B)(A), i.e. the degree to which A belongs
to B's fuzzy power set. Degree of subsethood will used to represent fuzzy causality.
Call the fuzzy-subset class .9. c F(2 x) a quantity space on X if every A c .~ can be
represented as A = Q w - Q for some Q, - Q e F(2X). Again " ~ " stands for abstract
negation or local complementation. Call some fuzzy-subset class d~ = F(2 x) a modifier
space on X if X e M. (X is the default modifier.) Then call the class ~ c F(2 x) a concept
space on X if ~' = .~ n M, i.e., if cr = {(Q w - Q ) n M: Q u - Q e ~, M c M}. Say that the
concept space ~' is causal if, for all C~, Cj c ~,
(1) Q, n M, c Qj n M j ~ - Q i n M, c - Q j n Mj,
(2) Q i n M i c - Q ~ c a M j ~ - Q i n M i c Q j n M j , for some Q i u ~ Q i , Q / u ~ Q j ~
and some Mi, Mj ~ dL
Then call the abstract pair ~ = (X, c~) a fuzzy cognitive space on X if c~ is causal (and
contains fuzzy sets).
qg contains the fuzzy nodes of the abstract FCM (fuzzy causal graph). The graph is
specified when the causally connected subsets of c~ • ~ are specified. This amounts to
defining a fuzzy edge function. (The edge function is fuzzy if its range set contains
more than two objects.) Formally, e: ~ • ~--> P is a fuzzy causal edge function on c~ if
e~ = ( C , Cj) = m~<2%(C~), the fuzzy-set membership of concept C~ in concept C f s fuzzy
power set, i.e. the degree of subsethood of C~ in Cj. The range set P can be any partially
ordered set; classically, P = [ 0 , 1]. To generalize unit-interval conventions, assume
e ( C , C~) < p for all p c P. So if P is the unit interval, e, = 0 and degenerate cycles are
prohibited. Then, finally, the abstract pair (e, ~:) is a fuzzy cognitive map on X if the
fuzzy causal graph (e, ~) is cycle-free.
72 a. KOSKO

5. Fuzzy causal algebra

A fuzzy causal algebra governs causal propagation and causal combination on a FCM.
It thus governs forward and backward chaining on a FCM. The algebra developed
below depends only on the partial ordering on P, the range set of the fuzzy causal
edge function e, and on general fuzzy-graph properties (e.g. path connections). The
algebra extends to any digraph knowledge-representation scheme.
Axelrod speaks of indirect and total causal effects on cognitive maps. Fix some
causal path from concept node C~ to concept node Cj, say Ci--' C k , - + ' ' " ~ Ck, ~ Cj,
which can be denoted with ordered indices as (i, k ~ , . . . , k n , j ) . Then the indirect effect
from C, to Cj is the causality Ci imparts to Cj (the degree of subsethood of C~ in Cj)
via the path (i, k ~ , . . . , k , , j ) . T h e total effect of C~ on Cj is all the indirect-effect
causality C~ imparts to Cj. Hence if there is only one causal path from C~ to Cj, the
total effect of C~ on Cj reduces to the indirect effect.
The operations of indirect and total effect correspond naively to multiplication and
addition of real numbers (field elements). Axelrod (1976) employs a causal calculus
of signs (+ and - ) that exploits this correspondence. The indirect effect of a path from
C~ to Cj is negative if the number of negative causal edges in the path is odd, positive
if the number is even. The total effect of C~ on Cj is negative if all indirect effects on
C~ on Cj are negative, positive if they are all positive, indeterminate otherwise. Hence
indeterminacy tends to dominate in this sign scheme. It can be removed, for a price,
with a numeric weighting scheme. If the causal edges are weighted with positive or
negative real numbers wu--i.e, if w~ = e~j e ~----as might occur if concept A causes B
five times as much as C causes B, then the indirect effect of C~ on Cj on path
(i, k~ . . . . , k n , j ) is the product Wik, X Wk,2X 9 9 "Wk.j and the total effect is the sum of the
path products. This weighting scheme generalizes the sign calculus, removes indeter-
minacy from the total-effect operation, and allows (requires !) finer causal discrimination
between concepts. But the real-valued requirement of P makes knowledge acquisition
difficult: forced numbers from insufficient decision information, different numbers
from the same knowledge source on different days, etc. The difficulties increase with
the number of knowledge sources and with the amount of knowledge-source
responses--the knowledge acquisition/processing tradeoff in this context.
A fuzzy causal algebra, and hence an FCM, bypasses the knowledge acquisition/pro-
cessing tradeoff. It allows fuzzy inputs to be processed as systematically as real-valued
inputs. The only price paid is a fuzzy output!
A fuzzy causal algebra is created by abstracting operations from multiplication and
addition that are defined on a (fuzzily) partially ordered set P of causal values. Let cr
be a causal concept space (on some underlying nonempty set X) and let e: c~ x c~., p
be a fuzzy causal edge function. (A really fuzzy e maps into fuzzy subsets of some
class P fuzzily partially ordered by fuzzy set inclusion (degree of subsethood).) Then
the simplest abstract operations are got by interpreting the indirect-effect operator I
as some minimum (infimum) operator and the total-effect operator T as some maximum
(supremum) operator--these operators depending only on P's partial order--and the
simplest of these operators are the rain (inf) and the max (sup). Formally, let there
be m-many causal paths from C, to Cj: (i, kit, k~, . . . , k ,t, , j ). for 1 -< l_< m. Let It(C~, Ci)
denote the indirect effect of concept Ci on concept Cj on the /th causal path. Let
T ( C , Cj) denote the total effect of C~ on Cj over all m causal paths. Then
I t ( C , , C j ) = m i n { e ( C p , C p + t ) : ( p , p + l ) ~ ( i , k,,.
~ .. , k ~,,,j)),
"
FUZZY COGNITIVE MAPS 73

T(Ci, Cj) = max I~(C. Cj),

I<l<_m

where p and p + 1 are contiguous left-to-right path indices.

Hence the indirect-effect amounts to specifying the weakest causal link in a path
and the total-effect operation amounts to specifying the strongest of the weakest links.
For example, suppose the causal values are given by P = {none ~<some ~<much<~ a
lot} and the FCM is given by

c~ ) c2
Some 1
Much A10t"
~':4
c3
Some
Alot ~ ~me
q

The three causal paths from C1 to C5 are (1, 3, 5), (1, 3, 4, 5), (1, 2, 4, 5). So the three
indirect effects of CI on C5 are
I1(C1, C5) = rain (el3, e35} -- min (much, a lot}
= much,
I2(C1, C5) = some, I3(C~, C5) = some. Thus the total effect of C~ on C5 is
T(C~, C5) = max {I~(C~, Cs), I2(C1, C5), I3(C~, C5)}
= max {much, some} = much.
In words, C~ imparts much causality to C5.
Finally, in fullest generality, I and T can be any respective t-norm (triangular-norm)
operator t and t-conorm s (see Klement, 1981; Yager, 1981, for definitions and
properties). If P is the unit interval, then the two triangular norms are related by
t(x,y)=l-s(1-x,l-y),
and by similar transformations for more general range sets P. Yager (1981) proves the
following maximality/minimality property of min and max:
t(x, y) <- min (x, y) -< max (x, y) <- s(x, y).
Thus t-norms other than min are never more causally lenient than rain and t-conorms
other than max are never more causally stringent than max. This property justifies
selecting min and max as default I and T operators when little or nothing is known
about how leniently I or how stringently T should causally behave.

References
AXELROD R. (1976). Structure of Decision: the Cognitive Maps of Political Elites. Princeton, N J:
Princeton University Press.
BANDLER, W. & KOHOUT, L. (1980). Fuzzy power sets and Fuzzy implication operators. In
Fuzzy Sets and Systems, vol 4, pp. 13-30.
74 B. KOSKO

KLEMENT,E. P. (1981). Operations on fuzzy sets and fuzzy numbers related to triangular norms.
11th International Conference on Multi-valued Logic. University of Oklahoma.
YAGER, R. R. (1981). Comments on "Textured Sets. An Approach to Aggregation Problems
with Multiple Concerns". IEEE Transactions on Systems, Man, and Cybernetics, vol 11, pp.
730-731.

Appendix: FCM matrix properties

Let C1, C 2 , . . . , C. be causal concepts and let e,i = e ( C , Cj) be the causal edge function
value, the amount of causality C~ imparts to Cj. For a cognitive map, e# =0, 1, or - 1
(0 represents an absence of a causal relationship). Let E = (eo)~,j~, represent the
matrix of causal edge values for the given FCM.
Suppose the FCM is a cognitive map. Then E is an adjacency matrix. It lists all
one-edge paths on the cognitive map. E 2 =[elj2)] = E x E lists all two-edge paths on the
cognitive map. For

e~ )= ~ e~kxe~
k=l

is nonzero only if there is a k' such that e~k, and ek'S are nonzero. Similarly
E 3, E 4 , . . . , E "-I list the effect of summing all three-edge, four-edge, . . . , ( n - D-edge
indirect effects. (Since FCMs are acyclic, there are no paths with more than n - 1
edges.) Then the total-effect matrix T is the sum of the powered matrices Ei:
n--I
T = Y . E'.

(A useful criterion for the existence of cycles in E is the following: E is acyclic if and
only if T's main diagonal is everywhere zero (otherwise, some concept is affecting
itself).) T is more generally known as the reachability matrix. If the above process is
repeated with E replaced E, the matrix of absolute values of E, then e# -(k) is nonzero
if and only if there are E~k)-many k-edge paths from C~ to C~. Such information is
useful when searching for forward and backward chains.
The conceptual centrality of causal concept node C~ is denoted by CEN(C~) and
defined by
CEN(C,) = IN(C,) + OUT(C,),
where

IN(Ci) = ~ eik,
k=l

O U T ( C i ) = ~ ekt.
k=l

The column sum of absolute values IN(C,) represents the number of concepts
causally impinging on concept Ci. Similarly the row sum OUT (Ci) represents the
number of concepts concept C~ causally impinges on. Hence the conceptual centrality
CEN (C~) represents the importance of concept node C, to the causal flow on the
cognitive map.
These concepts can be fuzzified and applied to more general FCMs. Suppose the
causal edge function is classically fuzzy, i.e. suppose it takes values in the unit interval.
FUZZY C O G N I T I V E MAPS 75

Then construction of fuzzy adjacency and fuzzy reachability matrices proceeds as

above by replacing eij with a causal indicator function (which is 1 if and only if eij> 0).
The fuzzy conceptual centrality of concept C~ is computed directly as in the non-fuzzy
case. The advantage now is that causal quality counts. A concept node can be connected
to fewer nodes than another concept yet still have greater conceptual centrality if its
connections are more heavily weighted.
When e maps into more general partially ordered sets, indicator functions can still
be used to compute adjacency and reachability matrices, since only cardinality of
connections matters. Causal conceptual centrality can be computed with the above
identities by replacing addition with maximum (supremum).