Professional Documents
Culture Documents
Subjective human expert knowledge in the form of Although AI was born as a concept of data processing
'rules of thumb' is still widely used and invaluable in that would mimic human problem-solving behavior,
food processing applications. However, expert knowl- McCuUoch and Pitts had already presented the first
edge cannot be easily dealt with by analytical and stat- simple binary model for a single brain neuron in the
istical methods commonly employed in the description early 1940s, leading Hebb to suggest in 1949 that infor-
of bioprocesses. Furthermore, such conventional methods mation in a neural network is stored in the 'weights'
are ill suited to dealing with the uncertainties character- (levels of importance) of the synapses 9. Neural network
istic of complex biological systems ~. In food process programs were initially developed as models of the
control, typically only a few process variables are avail- brain, but it was not until the publication of the super-
able for on-line measurement owing to the lack of suit- vised backpropagation learning procedure in 1986 by
able sensors. Even then the information obtained may Rumelhart et al.~ that the great potential of neural net-
be noisy, incomplete, scrambled, uncertain, erroneous work programs in modeling dynamic, nonlinear systems
or, in other words, 'fuzzy'. In such situations human was realized 9. Neural networks, like fuzzy systems, map
decision-making is also characterized by uncertainty inputs to outputs, but the beauty of a supervised neural
and subjectivity. Novel tools of 'artificial intelligence' network is that it learns from examples through
(AI; a paradigm first introduced by John McCarthy in iteration, without requiring a priori knowledge of the
1956), such as chaos theory, expert (knowledge-based) relationships between the process variables under in-
systems, fractal analysis, fuzzy logic, genetic algor- vestigation ~I. After the foundations of neuroengineering
ithms, naive (qualitative) physics, (artificial) neural net- were established, neural networks quickly became one of
works, and object-oriented programming, and their use the hottest resc:arch areas within AI. Of the vast literature
in combinations as hybrid systems, provide entirely new on the subject only a few books are referred to here ~2-~4.
opportunities in food science and technology. Although A discussion of the basic principles of fuzzy systems
the food industry has much to gain from the use of these is given first, with the underlying fuzzy mathematics
tools it generally lags behind the chemical and forest described in terms of hypothetical extrusion control
industries in the exploitation of modern computer appli- variables, with reference to Fig. l, as an example. Symbols
cations 2. In the following discussion, only fuzzy logic used are briefly defined in Box I. This is followed by a
and neural networks are covered in greater detail. (non-mathematical) description of selected, illustrative
Fuzzy (continuous) logic was introduced by Lofti food-related application examples. A brief account of
Zadeh 3 in 1965. The theory of fuzzy sets provides a con- backpropagation multilayered neural networks applied
venient means of dealing with uncertainties and trans- to recent bioprocess applications is then given; finally
ferring expert knowledge into quantitative functions that reference is made to the potential of hybrid systems.
can be processed by computer. No complex mathematical
relationships are required in the construction of fuzzy An introduction to fuzzy modeling
logic applications. Fuzzy models can be written in the form In the conventional theory of sets, each element x~
of easy-to-understand linguistic rules, and a number of (e.g. a measured extruder feed rate) of a set X, called the
commercial development tools are available 4. Of the 'universe of discourse' (e.g. the set of all possible feed
large number of books and papers published on fuzzy logic rates) definitely either belongs (e) or does not belong
and its applications only a few can be mentioned here ~-8. (~) to a defined subset A (e.g. the subset of feed rates
greater than 60 kg/h), such that:
T. Eerikliinen,P. Linko, S. Linko, T. Siimes and Y-H. Zhu are with the /IA(x~)= l if and only if xi e A (!)
Laboratory of Biotechnology and Food Engineering, Helsinki University of
Technology, 02150 Espoo, Finland. la,~(xi) = 0 if and only if x, ~ A (2)
Trends in Food Science & Technology August 1993 IVol. 41 o1993, ElsevierSciencePublishersLtd,{UK) 0924.2244/93/$06.00 237
Feed rate Feed moisture Change of watedeed
Rule 1: ~ 1
'Answer'
0 i
t
i
Xr Xm
Fig. 1
Graphical illustrationof membershipfunctions,and the use of defuzzification by taking the center of gravity of the fuzzy answer, to obtain a
crisp answer(y) for control action in the fuzzy control of flatbreadextrusion usingtwo fuzzy rules. See text for details.
In other words, the 'char~cteristic function'/J~(.~-,) is A,. A measured feed rate that 'rules o1" thumb' would
defined as I if .~'~ is a member ot'A (is greater than term 'a bit too high' for sorne operational control reason
60 kgth) and is defined as 0 ifxi is not a member ol'A (is might have a degree of membership o1", say, 0.6 in A I
less than 60 kg/h); Using conventional set theory to, (the set of high Iced rates), 0 in A.~ (the set of low feed
say, control extruder settings, only precise rules could rates) and 0.4 in a third set, A~ (the set of 'normal' feed
be applied, such as 'if P,~(.~'i) = I then turn the t'eed rate rates).
to 50 kg/h'.
Zadeh 3 generalized the characteristic function ,u..~(.~~)
Fuzzification
to a 'membership function': An important feature in fuzzy modeling is the 'fuzzifi-
p,~(.~,) e I0, ! I V.r, e A (3) cation' or mapping of numerical, crisp input data (e.g.
measured feed rate) into degrees of membership in sym-
describing the degree, within the closed interval 10,11, bolic, linguistic fuzzy sets such as 'high', 'low' and
to which each element .~-i in the universe of discourse 'normal'. The generalization of basic set theory allows
belongs to a 'fuzzy' subset, A. The degree of member- the mathematical expression and manipulation of im-
ship 0 means that the element definitely does not belong precise or 'fuzzy" rules tbr dealing with such input data,
to A, while the degree of membership I indicates that allowing the mathematical application of the 'rules of
the element, definitely does belong to A. In all other thumb" (e.g. 'if the feed moisture is too high then turn
cases the value of the membership function is a real the waterfeed down a bit') that constitute much human
number between 0 and I. For example, the set X of all expertise.
possible feed rates may be divided into several fuzzy The relationship between possible values .~i of a par-
subsets: fuzzy subset A~ might be the subset of 'high' ameter such as X and their degrees of membership in
feed rates and fuzzy subset A~ might be the subset of given subsets such as 'low' and 'normal' is usually
'low' feed rates. The maximum feed rate definitely represented by simple triangular or trapezoidal graphs
belongs to subset A~ and thus has a 'degree of member- (although more complicated nonlinear membership
ship' ! in AI; similarly it definitely does not belong to functions are sometimes employed): for example, see
subset At and thus has a degree of membership 0 in the leftmost pair of graphs in Fig. I. The degree of
238
Trends in Food Science& TechnologyAugust 1993 IVol. 41
Box 1. Basic set theory symbols