Behavioural
Processes,
0 1990 Elsevier
BEPROC
22 (1990)
197
197-212
Science Publishers
B.V. 0376-6357/90/$03.50
00323
Rule induction
’ Groupe
from examples for expert
systems
in mouse behavior
Jean-Arcady
Meyer
de Biolnformatique,
URA
‘and
AgnPs Guillot ENS,
686 CNRS-
’ Universith
I,*
46 rue d’lJlm,
75230
Paris,
France;
Paris X, France
(Accepted
8 August
1990)
Abstract
Quinlan’s observed preceding systems that
algorithm
ones
the
a given
they
of
rather
in this
study,
effective
the of
various order
in the cases
during
applied
out
knowledge
the physiolo&l
method
employed
Key words;
of the
here
Behavioral
induce, rules
rules
were
other
by
has
These
results
sequence;
Expert
the
other
system;
methods
Rule
by
order
activities
effective
and fails
two
It with
in is
means the
singled
out
predicting
the
moderately
respect
to drinking
light
activities.
having
no
during
only
in the
various
expert
It is shown
is
first
nest.
of these
into
mouse the
are discussed
determinants with
input
the seven very
in
sequences to the three
sequences.
the of
proven
and nest-building
is compared
then
determinism
grooming
behavioral
each activity
behavioral
exhibited
here
nest.
from
linking
the day. Among
and
of feeding
and to grooming about
on
activities
locomotion,
to
These
a Markovian
method
rest,
used
conditions,
be validated
displays
and of the second
occurrence
been
sequence.
might
succession but
has
in day and night
in
so that
random, night
ID3
on mice
similar
of current Lastly,
the
uses.
induction
Introduction The
use
diverse
as
of
expert
systems
mine
prospecting
Buchanan,
1974),
medical
molecular
biology
(Stefik,
has
become
(Duda
et
diagnosis 1981a,
al.,
widespread 1979),
(Shortliffe,
1981 b).
These
in
chemical
1976) systems
or
fields
of
analysis
experiment are actually
application (Michie scheduling computer
as and in pro-
198 grams
aimed
carrying TOR
at
is said
and
the
(Dreyfus
and
Whatever
who
system
elaboration
or
acquisition
turn
1983;
programs
operate
established
human
specialist
in
the
and
Rada,
induction
agglomeration
examples
themselves”
(Quinlan,
induction
sort
illustrating
the
activities
composing
these
these
most
probable
both
day and
other
latter
rules
mice
were
night
interpreted
effective
at
PROSPEC-
(Forsyth,
1984)
laboratories
subject
to
question
In
of the
rules
order
its own 1989).
to
rules
to
the
its
both
“knowledge disadvantages,
render
these
and the computer
stage,
the
it automatic,
of operation
Among
of which
This as
avoid
sought
the
the of the
it applies,
between
known
by their
reproduce
effectiveness
involved.
be
have
to
the
collaboration to
supplied
attempt
(Michalski
efforts,
“structures
et
most
have
underlying
are discovered
through
out
103
and
analysis
of the
1979). been
From
rules
sequences use
carried
actual
which
conditions.
the Lastly,
behavioral
two
into
of
the
identify
systems
a given
these
expert
sequences,
and
1982), in
the
rules
intended
Systems
different this
in question. to predict
sequence,
were
rules
of
the
under
validated
they
a
mice,
implementing
behavioral
the
(Quinlan, observed
succession
a program
to
expert
course
with
sequences
govern
input
examples
into
in
here
behavioral
were
of these
incorporated
other
an
quality
of examples
has
activity
in
application
by itself
unknown
made
future
and
of rules
1977).
by means
algorithm.
examples
Next
basis
come
Forsyth,
process
of this
This
on the
intelligence
1986;
a chaotic
of
has
to acquire
the
algorithm.
are
the
field
artificial
explaining
particular
as
chemistry
claims
a lengthy
(Feigenbaum, in
a program
application
in
In consequence,
upon
entails
addressed
An
is
system
deposit
use
these
them
experts.
inevitable,
efforts
Forsyth
that
the expert
common
although
dependent
bottleneck”
is to allow
Thus
a molybdenum
in
1983)
has
more
seemingly
research
in discovering
a program
the
and
several
in
is strongly
and
process
expert.
1989).
of such
sensitive
reasoning
reportedly
McCorduck,
of one
scientist
automatic
is
the case, these
reasoning expert
an
as that of a human
DENDRAL
and Dreyfus,
programmer,
al.,
tasks
to have succeeded
System
(Feigenbaum
that
producing
out certain
against
evidenced
were
and discussed.
The ID3 algorithm The
103
algorithm
set of properties may
be either
rules
that
other
symbolic
enable
words,
is applied
or attributes to
predict
is
used
to
the various
rules
the
production list
concepts
form
rules
or numerical,
one
103
describing take
to a series
of
(Klahr
of examples,
and illustrating
either
and the
which
induce, of which
each characterized
particular
object
concept from
is
of the
and are created
(Breiman
et
al.,
and processed
attributes
is to discover
by which the
are specific
by a given
These
algorithm
examples,
examples
tree
concept.
illustrated
individual
these
a regression
et al., 1987)
one
example. general
instances. 1984)
or
In rules
These a set
in the computer
of in
form. In this
illustrating
way,
from
2 different
the 10 following concepts:
examples,
each characterized
by 3 attributes
and
199 CONCEPTS
ATTRIBUTES RACE
FUR
NOISEMANIFESTATIONS
Siamese Abyssin
ruffled-up normal normal
Miaow Purr Miaow
normal
Miaow Shhh
Siamese European Siamese European
ruffled-up normal
European
normal
European Abyssin
ruffled-up normal
Siamese
normal
algorithm
Purr Shhh Shhh Miaow Purr
103 induces the regression
tree:
FUR ruffled-upI CONCEPT
+
normal I NOISE-MANIFESTATIONS
h4iaowA
Shhh
CONCEPT-
CONCEPT-
corresponding to the equivalent list: (FUR (ruffled-up (CONCEPT + )) (normal (NOISE-MANIFESTATIONS
or the following 1 23 4 -
production
If FUR = ruffled-up THEN CONCEPT = + IF FUR = normal AND THEN IF FUR THEN IF FUR THEN
CONCEPT = normal CONCEPT = normal CONCEPT
= + AND = AND = +
CONCEPT
+
(Miaow(CONCEPT (Purr (CONCEPT
- )) - ))
(Shhh
+ )))))
(CONCEPT
rules:
IF NOISE-MANIFESTATIONS
= Miaow
If NOISE-MANIFESTATIONS
= Purr
If NOISE-MANIFESTATIONS
= Shhh
Once formulated, these rules can be used to predict which concepts are illustrated by new examples, solely on the basis of their attributes. If the concept that is actually
200
Illustrat\on of linear and non-linear separabilities. In case A the examples,
Fig. I,
rharacterired
by two numertcal attributes,
decrde, followrng sentative Positive
point on a fictitious values
simple
straight
not linearly examples,
characterize
computation
separable:
the
> Xn AND
+
examples,
more than a straight +,
Xj < Xm)
AND
line
i should
be assigned
on a surface.
is required
supposedly
For instance,
one may
f of the corresponding
and negative values,
different.
a production
separable.
projection
the two categories
they are obviously
to Concept
yielding
variable F, that an example
line OZ separating
although
be assigned tf (Xj
a numerical
X and Y, are linearly
the
-
repre-
to Concept
examples,
+.
with
In case 5, the examples
to separate
the
In order to ascertain whether
+
from
example
a are
the
-
j should
rule such as
IF (Yj > Yn AND
Yj < Ym)
THEN
CONCEPT
= -t
must be called upon
instanced by each of these examples is known,
a comparison
reality makes it possible to validate the rules in question. In the application just mentioned the RACE attribute
between prediction
contributes
and
no information
about the concepts illustrated and accordingly need not be made to appear in the induced rules. In more complicated cases, one may seek which attributes should be retained and in what order. This query is answered by t53 on the basis of an evaluation of the discriminating potential of each attribute in terms of entropy, to unnecessary complexities in the regression tree generated. Accordingly, the attributes are introduced into this tree one at a time in decreasing-entropy order, an approach which entails integrating first those attributes that yield the most informa-
avoid
tion about the concepts under consideration.
The process is terminated
as soon as the
tree generated permits an accurate classification of all available examples or, in other words, whenever all the rules induced summarize perfectly all information provided about the concepts in question. Thus conceived, algorithm 103
analogies with certain more discriminant analysis” (Jenrich, 1977). In comparison with these latter, however, it offers the advantage of allowing both qualitative and quantitative attributes to be used and especially allows the management of non-linear separabilitites, as illustrated in Figure 1. Furthermore, it produces rules that are much more intelligible than the weights associated with the
classical
data-analysis
factors habitually ID3 may likewise to animal and
displays
obvious
such as the “stepwise
extracted in multivariable analyses. The rules induced by algorithm be compared to the grammar rules which have already been applied
behavior
Rosebrugh,
methods,
(Kalmus,
1979).
1969;
If these
Fentress
latter
and Stilwell,
can also
1973;
be inferred
Westman,
from
examples,
1977;
Rodger
it appears
201
that the sole methods actually capable of doing so (Gonzales and Thomason, 1978) yield only context-free syntactic rules, whereas the rules inferred by ID3 can be context-dependent. Thus the information FUR = ruffled-up triggers CONCEPT = + independently of any context, whereas the information can only lead to a given decision in function of the sound context imbedded. Designed most serious
the decision FUR = normal in which it is
to operate with non-redundant and non-contradictory examples, the drawback to the ID3 algorithm is its background noise sensitivity, that is
its susceptibility
to errors
committed
in the description
of the examples
given to it.
Such errors may be conveyed by one or more of the attributes, or even by the concepts that are illustrated. Their most common impact is to produce complex regression trees, the last branches of which being appended in order to account for rarely appearing cases, consequently
of dubious
statistical
significance
(Forsyth
and
Rada, 1986; Hart, 1989). A number of modifications to algorithm ID3 have been considered. Some aim at consolidating its statistical basis (Quinlan, 1985; Cestnik et al., 1987; Clark and Niblett, 1987), while
others seek to generate the simplest
al., 1988; Fisher
and Schlimmer,
1988; Quinlan,
possible
regression
1988). Yet others
trees (Cheng et
make it possible
not
to process all the available examples in one batch, and thus to carry out what is called jncrement~l learning (Schlimmer and Fisher, 1986; Shapiro, 1987; Utgoff, 1988, Arciszewski and Mustafa, 1989). The majority of these methods present computation time or memory capacity obstacles; what is more, their very multiplicity poses a problem of choice. This
is why algorithm
ID3 has been used here in its original
being applied to a limited number of examples once non-redundant and non-contradictory.
Experimental
which
have been selected
form,
to be at
Data
Algorithm ID3 has been used to ascertain whether rules exist that enable one-once he knows the nature of the three activities (Act-3, Act-2, Act-l) preceding a fourth one (NEXT-ACT) within the behavioral sequence of a mouse-to predict the nature of this fourth activity. To accomplish this, the behavioral sequences of ten mice were placed under continual observation during eleven and one half consecutive hours, in day condition. Ten other mice were observed over an equivalent period of time, but in night condition. Each mouse was kept isolated in a 23 x 8 X 8 cm transparent polystyrene cage containing sawdust litter, water, and food ad libitum, in addition to a cotton ball by way of nest-building material. This cage was, in turn, set in a larger enclosure, 52 X 125 X 90 cm, with transparent PVC partitions; this could be lighted either by white fluorescent tubes (IO0 lux)-to represent day condition-or by red ones (10 lux)-for night condition. The temperature in this enclosure was held between 19 and 21*C. and its humidity between 60 and 70%. In the course of these observations ten different activities were identified (Guillot, 1988) and their successive
occurrences
noted down. These
acts are respectively:
202
ACTS
CODE
Rest Sniffing Locomotion Feeding
R
Nest-building Drinking
N D
Grooming Grooming
out of the nest during rest
GA CR
Grooming Grooming
before activity after activity
GB GF
S L F
Owing both to memory capacity and computational time considerations and to the fact that results obtained here were to be compared to others, the present study was restricted to seven activities, with sniffing being omitted and the three categories grooming in the nest (GR, GB and GF) being recombined into one (GN). For each the two lighting conditions, five of the behavioral sequences thus obtained served a learning batch, with the five remaining ones as a validation batch. The lengths these sequences,
reckoned
in terms
of the number
noted, varied between 159 and 243 for the diurnal the nocturnal one.
of successive
activities
of of as of
that were
series and between 331 and 457 for
The examples input into the ID3 algorithm were generated in several steps. First, all the subsequences, four consecutive activities long, were excerpted from the behavioral sequences of each of the twenty mice under observation. Thus, a sequence beginning with F-L-D-GA-R-L yielded the following successive examples: CONCEPT
ATTRIBUTES Act-3
Act-2
Act-l
NEXT-ACT GA R
F
L
D
L
D GA
GA R
D
L
Under these conditions, the diurnal and noctural learning sequences furnished 990 and 1916 examples respectively, while the corresponding validation sequences furnished 818 and 2075. The examples drawn from the learning sequences were then filtered so as to retain only one among all those examples that would prove redundant because of being characterized by the same attributes and being illustrative of the same concept. A second cull was effected for the purpose of detecting contradictory examples, that is, those characterized by the same attributes but illustrating different concepts. Rather than eliminate all these examples, it was decided to retain those embodying a situation both representative and predominant. Thus, where 12 examples had been characterized by the same 3 attributes, with 5 of them illustrating concept R, 4 concept L, and 3 concept N, this set of examples would have been taken contradic-
203 tory.
Since the majority
grouping,
5, is higher
neither
3 (the arbitrary
threshold
of
representation) nor the sum of the other groupings (in this case 4 + 3), all 12 examples would have been eliminated. However, had these examples been distributed as 8 R’s, 2 L’s, and 2 N’s, they would have been replaced by a single example associated with R. Lastly, a preliminary study (Guillot, 1988) having revealed difficulties in predicting, with such an approach, the activities drinking and grooming out of the nest, the examples illustrating the corresponding concepts have been eliminated. These two activities have nonetheless been retained as possible attributes. At the output end of this filtering process, the diurnal learning batch was composed of 33 examples
and the nocturnal
one of 26. The validation
other hand, were only screened for examples
of the nest. Reduced respectively faithful
reflection
of the original
concerning
batches, on the
dr;nk;ng and grooming
out
to 638 and 1421 examples, they continue to be a behavioral sequences, with their redundancies and
their contradictions.
Results When applied to the diurnal batch of learning examples, algorithm ID3 gave rise to the regression tree coded in the ensuing list for which the most discriminating attributes turn out to be Act-2, followed by Act-l: (Act-2
(feeding
(locomotion (nest
(Act-l
(Act-l
(groom-out (locomotion (drinking (groom-out (locomotion ~groom-in
(groom-in
(Act-7
(nest (locomotion (rest
(rest
(NEXT-ACT (NEXT-ACT
(FEEDING (FEEDING
) >) )) )
(NEXT-ACT (NEXT-ACT (NEXT-ACT
(LOCOMOTION (LOCOMOTION (NEST ) ) )
(NEXT-ACT
(NEST
(NEXT-ACT (NEXT-ACT (NEXT-ACT
(REST ) ) ) ) ) (GROOM-IN )) ) (FEEDING ) ) )
(NEXT-ACT (NEXT-ACT
(GROOM-IN (REST ) ) ) )
>) ) ) ) ) ) )
) ) )
) )) ) )
Eleven rules are coded into this list and are expressed in more intelligible Table la. In this table, a rule such as F-GA --, F can be read to signify: I? Act-2 = Feeding AND Act-l
= Groom out of the nest CUFF
NEXT-ACT
form
in
= Feeding
or, “If feeding has been followed by grooming out of the nest, then the next activity will again be feeding.” Similarly, a rule such as L-* --+ L signifies that locomotion followed by any other activity tends to determine a new locomotion phase. The nocturnal learning batch, in like manner, gave rise to the derivation of the following coded list for which the order of excerption of the discriminating attributes is the inverse of the preceding:
204 TABLE
1
Day and night rules. la-Day
1 b-Night
Rules
No. 1
F-GA
2
+ F L+F
3
D+L
4
L-*+L
5
N-GA
6
-+ N
GN+R
8
CN-N
9
1
F+L
2
GA-,L
3
F-L
4
N-*+N
5
L-*N
7
--+ CN
+ F
D+L
6
N+L
7
CN+R
8
R-CN
L+F
10
R-CN
11
R-*-R
(Act-l
(feeding (groom-out (locomotion
(Act-2
(feeding
(nest (drinking
(LOCOMOTION (LOCOMOTION
(NEXT-ACT (NEXT-ACT (NEXT-ACT
(FEEDING
(groom-in (rest
(NEXT-ACT
8 rules
TABLE
(NEXT-ACT (NEXT-ACT
(NEXT-ACT (NEXT-ACT
(nest
The
rules
No.
of Table
lb
correspond
to this
) ) ) ) ) )
)) )
(NEST ) ) ) ) ) (LOCOMOTION (LOCOMOTION (REST ))) (GROOM-IN
) ) ) ) ) ) ))))
list.
2
Day validation
(PRED
figure 60, for example, the contrary, PRED >
= predicted behavior; OBS = observed indicates that the expert system
it incorrectly R
predicted L
L (locomotion) F
N
behavior,
correctly
2 times
! = indeterminations).The
predicted
R (rest) 60 times.
instead of R.
D
GA
GN
?
OBS v R
60
2
1
0
0
0
3
4
L
21
212
6
2
0
0
1
26
F
2
8
53
1
0
0
0
34
N
14
1
9
53
0
0
0
32
D
0
0
0
0
0
0
0
0
GA
0
0
0
0
0
0
0
0
GN
0
19
0
0
0
0
67
7
Number Ratio?/Nb
of examples = 638. examples = 0.16.
Ratio Success/(Success+ Ratio Success/(Success
Failures)
= 0.83.
+ Failures + ?) = 0.70
On
205 TABLE
3
Statistics Rules
on day
rules.
No.
Failures
Successes
% Successes
15
7
68.18
2
30
6
83.33
3
9
2
81.82
28
87.88
203
4 5
7
2
77.78
6
46
1
97.87
7
24
5
82.76
8
10
4
71.43
9
8
3
72.73
10
57
0
100.00
36
32
52.94
445
90
83.13
11 Total
These
series of rules were subsequently validated using the corresponding batches. This process entailed incorporating them into two simple expert
two
validation
systems that examined each example of a given batch in turn in order to ascertain-on the basis of the attributes Inherent thereto-to what concept it should be assigned, then to compare the result of this operation with the concept actually instanced by the example in question. In other words, these expert systems make it possible to scan the examples describing the validation sequences, to make a decision at any moment -in view of the nature of the three activities preceding it in a sequence-as to the nature of the fourth
ensuing
activity actually displayed is found diagnosis
TABLE
and lastly
to confront
this
prediction
In this process,
with
to correspond to a given set of attributes, the expert system renders a of indetermination. Once an entire validation batch has been scanned,
validation
PRED
>
R
L
F
N
D
GA
GN
?
v
R
12
L
8
868
F
0
44
N
1
14
7
41
0
D
0
0
0
0
0
GA
0
0
0
0
0
0
0
0
GN
0
11
0
0
0
0
13
0
Number Ratio?/Nb
the
if no rule
4
Night
OBS
activity,
by the mouse under examination.
5
of examples
=1421.
examples
= 0.18.
Ratio
Success/(Success+
Failures)
Ratio
Success/(Success+
Failures
0
0
0
0
0
0
0
0
0
4
12
0
0
0
206
0
0
50
0
0
0
125
= 0.91. + ?) = 0.74.
0 0
206 TABLE
5
Statistics on night rules Rules
No.
Successes
Failures
% Successes
1
234
5
2
261
35
88.18
3
125
7
94.70
4
41
12
77.36
5
285
21
93.14
6
88
13
87.13
7
12
9
57.14
8
13
4
76.47
1059
106
90.90
Total
several
valuable
Statistics
system
that was
run.
The
results
of these
and 5 as concerns
are collected validations
the success
about
are given
and failure
97.91
the successes in Tables
statistics
and failures
of the expert
2 and 4, completed
related
by Tables
3
to each rule.
Discussion The
results
described
observed
behavioral
validation
examples
90.9%
accurate
significantly
for
often
were
individual
rule,
(70%
to predict
the following merely
derived
It will classical
method
be noted Markovian
7, 8, and 10; night
to the
building
the
the
errors
are
relate
the and
furthermore basis
at .OOl).
rules
from on
day expert
sole
of
Finally,
grooming
the
some
in the nest,
7 and 8). Indeed,
and the occupation
57.1
that
use only
correlations respect
the
night
of the
these
nest
3
must
that
(Chatfield
this
the
validity
random.
of successes would
and
seem
Lemon,
causal 1970;
to
obtained
making
up a
are, on
at the very
number
activity-or
descriptive replaces
52.9
least,
of examples
to be of a relatively
of conclusion direct
results
successions
as 72%
to the
type
method
of each
from
activities
These
preceding
and not
pertinence ranging
of the
one
has a purely this
the
ones.
can be estimated
determinism
here
on
at validation
of different
are in no way
Naturally
applied
for
from
of the number this
analyses
to 97.9%
be drawn
induced one.
dependent
percentage
that the succession
to an extent ratio
reflect in this
is clearly
a success
mouse
Furthermore,
thus
they
may
as the rules the
procure on
when
appear
determined
the fact that
they
be expected
by definition,
behavior
of the
to the average
insofar
for
diagnostics
induced
few
accurate
significant
and from
it would
and 74%).
nature
commit
(Chi2
related
displays
sequence
contrary,
according
(83.1% to
automatically
they
of occurrence
(day rules
of conclusions
behavioral
The were
systems
Indeed
each other.
which
Foremost,
expert
conditions
that
valid
general
for the day rules
A variety
that valid.
one).
those
probability
follow
satisfactory
same
night
obviously
This
the
the
induced
being
logically
here.
the
and to nest-building
activities
100%
are
than
activity’s
of the rules to rest
prove
for
better
individual
above
sequences
must use,
two-in
local order
be connected and that
the
to rules
relationships. advantageously
Slater,
1973;
the
Bekoff,
more 1977;
207 TABLE Day
6 expert
PRED
on nocturnal
>
OBS
sequences
R
F
L
N
D
GA
GN
?
v
R
12
0
1
0
0
0
0
L
8
753
36
3
0
0
0
80
F
3
29
138
12
0
0
0
205
N
1
6
7
44
0
0
0
55
D
0
0
0
0
0
0
0
0
GA
0
0
0
0
0
0
0
0
GN
0
9
0
0
0
0
13
2
Number
4
of examples
Ratio?/Nb
= 1421.
examples
= 0.24.
Success/(Success
Ratio
+ Failures)
Ratio Success/(Success
Fagen
and
Young,
1978).
These
relate
to
a given
order,
are
determinism the same
(such only
would through
defined
the
1 2
which of
address
the totality
identifying
F) and a first
since
are strictly
appear
that
by results
however,
order
is slightly
at once
determinism
readily
when
and vice more
over
the
versa
complex
The
diurnal
corresponding
obtained
of rules.
of its rules
do not
identical the
system
number
certain
respectively)
of the sequence both (such
No.
and order
as F -+ L) within
former
night
(Tables
than
the
also
differs
(1, 5, and 11, concerning apply
and nocturnal
determinisms 6 through nocturnal
on day
rules
applied
to nocturnal
Successes
9). one
9
sequences
36
% Successes 20.00
7
9
68
11
86.08
4
685
37
94.87 42.86
94.70
5
3
4
6
41
12
77.36
7
7
1
87.50
8
1
0
100.00
9
4
0
100.00
10
12
0
100.00
11
5
11
31.25
1I19
88.97
960
The
are
diurnal
in that extent
nestbuilding,
at night.
Failures
close
examples
to a certain feeding,
situations,
present
validation
125
Total
a second
7
Statistics Rules
+
day expert
by a larger
the latter,
TABLE
rules
as evidenced
determinism,
rest,
two
nevertheless
analogies, run
as F-GA
latter, incapable
sequence.
Although it
= 0.89.
+ Failures + ?) = 0.68
it is from and
208 TABLE 8 Night
expert
on diurnal
PRED >
sequences
R
F
L
N
D
GA
?
GN
OBS v R
59
9
0
0
0
0
0
2
L
18
250
0
0
0
0
0
0
F
1
23
30
1
0
0
3
40
N
14
10
6
46
0
0
1
32
D
0
0
0
0
0
0
0
0
GA
0
0
0
0
0
0
0
0
GN
0
31
0
0
0
0
62
0
Number
of examples
= 638.
examples
= 0.12.
Ratio?/Nb Ratio
Success/(Success+
Ratio
Success/(Success
The
diurnal
considers while be
the latter
respect
the
processes it
processes
is
also relies
involves
with
mainly
less
heavily
processes the
of sequences.
activated
the
seems
more
in which
order
to the 2 series
which
= 0.79.
determinism
that the former
linked
order
Farlures)
+ Failures + ?) = 0.70.
by day
same
activity
that
are potential
triggers
for
local upon
than
of order
usually what
latter
Berridge
point
particularly
excerpted
with
transitions-in
another-and
et al. (1987)
one
that the second
behavioral
follows
if
processes,
must
were
be stressed
particular and
one,
Markovian
attributes
it should
imply
precedes
nocturnal
order
1. This
discriminating
Incidentally,
the
second
that
have termed
these
reciprocal
transitions. Moreover, elsewhere rules the
some on
7 and 8 describe beginning
Hazout
of
the
behavioral
results
obtained
sequences respectively
of an activity
in the
here mouse.
confirm Thus
the subsequences
period
that
were
brought
viewpoints
day rules
characteristic to
light
already
held
7 and 9 and
night
of the in Guillot
et al. (1989).
TABLE 9 Statistics Rules No.
on night
rules applied Successes
to diurnal
sequences Failures
% Successes
1
53
3
2
80
29
73.39
3
30
6
83.33
4
46
1
97.87
5
55
3
94.83
6
62
38
62.00
7
59
33
64.13
8
62
4
93.94
447
117
79.25
Total
94.64
end
and of
(1988)
and
209 Lastly, when examining individually each activity identified in this study, clearcut differences become evident. Rest and /ocomotion are very readily recognized by the expert systems in question, and the essential part of their determinism is accounted for by the rules
induced, if we judge by the corresponding
tion. Grooming other activities,
in the nest is fairly often confused with locomotion, nor does it often result in cases of indetermination.
low rates of indeterminabut not with Feeding and
nestbuilding are recognized for the most part by the expert systems, but these activities are prominent in that the corresponding indetermination levels are quite high. It is logical to conclude that their determinants information
conveyed by the activity or activities
correlates
immediately
only partially
with the
preceding them within
the behavioral sequence. Where feeding is concerned, this point is obviously to be brought into relation with what is already known about the determinism of this activity. It has in effect been demonstrated elsewhere (Guillot and Meyer, 1987) that the triggering and termination of food intakes in mice appear dependent upon the instantaneous energy flow entering the lean tissues, in conformity with a general hypothesis proposed by Booth (1978). The rules defined here governing feeding (day rules 1, 2, 9 and 3 for night) link this act with preceding locomotion and feeding periods. Now it will be recalled that locomotion and feeding are two energetically costly activities for the mouse (Meyer and Guillot, 1986) and that, in contrast with drinking, they last a relatively long time. This explains why the fact of knowing that these activities have just been engaged in contributes significant information as to the mouse’s net energy gain and accordingly about the importance of its internal energy flows. This information is nevertheless imprecise enough to explain a certain number of failures on the part of the corresponding tion that were observed,
rules. It may likewise explain the high rates of indeterminainsofar as many examples from the learning batches may
have been declared contradictory in terms of preceding activities, whereas they would not have been considered so, had they been described in terms of instantaneous energy flows. A similar type of reasoning may be called upon to explain the failure of this approach where drinking is concerned. The corresponding determinism is partially known in mice (Cuillot and Meyer, 1985; Guillot, 1988) and would appear to depend upon the water content of the intra- and extracellular compartments, in accordance with general mechanisms proposed by Toates (1974). It may be thought that no unequivocal correlation is to be found between the time course of these water contents and the succession of particular activities performed by the animal, and that therefore no rule enabling drinking to be predicted can be linked with preceding activities alone. The same may be true of grooming out of the nest, of which it is thought (McFarlane and Epstein, 1981) that they are in relation with insensible water loss and accordingly with the animal’s water balance.
Conclusion The rule induction method presented herein is one that can prove useful in describing and interpreting animal behavior. In particular, it is shown to have greater potential than other methods with a similar field of applicability. In comparison with classical data analysis,
it is not encumbered
by problems
of linear separability
or of the
210
intelligibility of the results obtained. In comparison niques, it is not limited by context problems. Lastly,
with grammar induction techin comparison with Markovian
approaches, it is not subject to the estimation of an overall order of dependence. Furthermore, it is capable of constituting an effective complement to the melody excerption method proposed elsewhere (Hazout et al., 1989). Applied to the same behavioral
sequences
subsequences than
random
consecutive
frequency
within
governed
by a quite
preceded
the one
Finally,
the
accordingly
in mice, these methods
several
local
lends
itself
relatively
subtle
and another.
to
comparative
this
melodies
at most
that
observed may
with
if repetitive
with
a better
nevertheless
the two
activities
be that
individual.
rules
that
one
can
be individually
validated.
it enables
to
since
between
reason,
indicate
indeed
such
applications,
existing
For
ethology
by the
identifies
both
are
correlating
practical
differences
in fact
long
sequences,
performed
method
condition like
these
determinism,
currently
present
activities
one
may
one
observational
well
envisage
or
It
pinpoint
experimental
applications
in areas
or psychopharmacology.
References Arciszewski,
T. and Mustafa,
learning Bekoff,
Principles
M. 1977.
behavioral
M. 1989.
and techniques
Quantitative
taxonomy
K.C.,
Fentress,
learning:
the user’s
perspective.
ed.). Chapman and Hall.
of three
areas of classical
and behavioral
animal behavior (Hazlett, Berridge,
studies
Inductive (Forsyth,
variability.
ethology:
In Quantitative
In Machine
social
methods
dominance,
in the study
of
action sequence
of
ed.). Academic Press.
J.C. and Parr, H. 1987.
Natural
syntax
rules
control
rats. Behav. Brain Res. 23, 59-68. Booth,
D.A.
models: Breiman,
1978.
Prediction
computable L., Friedman,
trees. Wadsworth Cestnik,
U.M.,
Chatfield,
behavior
of feeding control Olshen,
R.A.
from
energy
(BOOTH,
and Stone,
flows
in the
rat. In Hunger
ed.) Academic Press.
C.J. 1984.
Classification
and regression
international.
B., Kononenko,
sophisticated
of feeding
theory
I. and Bratko,
users.
In Progress
C. and Lemon,
I. 1987.
in Machine
R.E. 1970. Analysing
Assistant
86: a knowledge-elicitation
learning (Bratko sequences
tool
for
&Lavrac, eds). Sigma Press.
of behavioral
events.
J. Theor.
Biol.
29,
427-445. Cheng, J., Fayyad, U.M., version Clark,
of ID3.
P. and Niblett,
(Bratko Dreyfus,
H.L.
and Qian,
T. 1987.
Induction
and Dreyfus,
SE.
1989.
Z. 1988.
Improved
learning.
in noisy
Mind
in the era of the computer.
R., Cashnig,
Mineral Univ.
K.B.
decision
Morgan
domains.
trees:
a generalized
Kaufmann.
In Progress
in Machine
learning
and Lavrac, eds). Sigma Press.
expertise Duda,
Irani,
Proc. 5th Int. Conf. on Machine
J. and Hart,
Exploration.
over machine.
The
power
of human
intuition
and
Bell and Bain Ltd.
P. 1979. Model
In Expert systems
design
in the Prospector
in the micro-electronic
Consultant
age. (Michie,
System
for
ed). Edinburgh
Press.
Fagen, R.M.
and Young,
D.Y.
1978.
Temporal
patterns
of behaviors.
In Quantitative
Ethology
(Colgan, ed.). Wiley. Feigenbaum,
E.A. 1977. The art of artificial
edge engineering. Stanford
intelligence.
rep. STAN-CS-77-621.
1: Themes
Stanford
and case studies
Department
of computer
of knowlScience.
Univ.
Feigenbaum, Fentress,
Tech.
E.A. and MC Corduck,
J.C. and Stilwell,
224, 52-53.
R. 1983. The fifth
generation.
F.P. 1973. Grammar of a movement
Addison
sequence
Wesley.
in inbred mice. Nature.
211 Fisher,
D.H.
and Schlimmer,
J.C. 1988. Concept simplification
Int. Conf. on Machine learning. Forsyth,
R. 1984. Expert systems,
Forsyth,
R. 1989. Machine
Forsyth,
R. and Rada, R. 1986. Machine
retrieval.
R.C.
Addison
Principles
learning.
and case studies.
Principles
and Thomason,
A. 1988.
Contribution
descriptive,
Specialite
learning
M.G.
1978.
Un
a I’etude
modele
Syntactic
pattern
des sequences
Bull.
ecophysiologique
and information
recognition:
an introduction.
comportementales
These
de la souris:
ap-
de Doctorat
de I’Universite
Paris
7.
du temps et des ressources
sensorimotrices
chez la
du determinisme
des comportements
alimentaire
et
Ecol. 16, 1, 69-75.
A. and Meyer,
J.A. 1987. A test of the Booth
patterns of mice. Appetite. Hart, A. 1989. Machine S., Guillot,
energy flow
model (Mark
3) on feeding
8, 67-78.
induction
ing. In Machine learning. Hazout,
in expert systems
Biomathematiques.
dipsique. Guillot,
Chapman and Hall.
application
causale et fonctionnelle.
A. and Meyer, ].A. 1985. Allocation
souris.
Chapman and Hall.
and techniques.
Wesley.
proches Guillot,
accuracy. Proc. 5th.
Wiley.
Gonzalez, Cuillot,
and prediction
Morgan Kaufmann.
as a form of knowledge
Principles
A. and Meyer,
and techniques
J.A. 1989.
acquisition
(Forsyth,
Extraction
in knowledge
engineer-
ed). Chapman and Hall.
of melodies
in behavioural
sequences.
Behav. Proc. 20, 61-74. jenrich,
R.I. 1977.
(Enslein, Kalmus,
Stepwise
Ralston
discriminant
and Wilf,
H. 1969. Animal
analysis.
In Statistical
methods
for digital
computers
eds). Wiley.
behaviour
and theories
of games and of language. Anim.
Behav. 17,
67-617. Klahr,
D., Langley,
ment. The MIT McFarlane,
P. and Neches,
B.A. and Epstein,
A.N. 1981. Biobehavioral
the rat. Behav. and Neural Meyer,
J.A. and Guillot,
mouse.
R.S.,
Biol.
D.
Carbonell,
artificial
spectroscopy
of learning
and develop-
determinants
of evaporative water loss in
energetic
cost of various
behaviors
in the laboratory
T.M.
1983.
Machine
learning.
An artificial
intelli-
Verlag. B.C.
1974.
intelligence
(Carrington,
Current
to the
status
of the
interpretation
of
heuristic
mass
Dendral
spectra.
In
Program
Computers
for for
ed). Adam Hilger.
Quinlan,
J.R. 1979. Induction
Quinlan,
J.R. 1982. Semi-autonomous
readings in expert systems Quinlan,
models
83, 3, 533-538.
J.G. and Mitchell,
and Buchanan,
applying
The
Physiol.
gence approach. Springer Michie,
system
33, 101-116.
A. 1986.
Comp. Biochem.
Michalski,
R. 1987. Production
Press.
J.R. 1985. Induction
over large data bases. Tech. acquisition
(Michie,
ed.). Gordon
of decision
Rep. HPP-79-14.
of pattern-based
Stanford
knowledge.
University.
In Introductory
and Breach.
trees. Tech.
Rep. 85-6.
New South
Wales
Institute
of
Technology. Quinlan,
J.R. 1988. Simplifying
systems Rodger,
(Gaines
R.S. and Rosebrugh,
acts. Anim. Schlimmer,
decision
trees.
In Knowledge-acquisition
for knowledge-based
and Boose, eds.). Academic Press. R.D.
1979.
Computing
a grammar
for sequences
of behavioural
Behav. 27, 737-749.
J.C. and Fisher,
D.H.
1986. A case study of incremental
concept induction.
Proc. 5th
Int. Conf. on Al. Morgan Kaufmann. Shapiro,
A.D. 1987. Structured
Shortliffe, Slater,
P.J.B. 1973. Describing
Klopfer Stefik, Stefik, 141-I
induction
E.H. 1976. Computer-based ed.). Plenum
70.
1981b.
sequences
of behaviour.
Turing
Institute
Mycin.
In Perspectives
Press.
Elsevier. in Ethology
(Bateson
and
Press.
M. 1981a. Planning M.
in expert systems.
medical consultations:
Planning
with constraints
(Molgen:
and meta-planning
Part 1). Artificial
(Molgen:
Part
2).
Intelligence. Artificial
16, 111-140.
Intelligence.
16,
212 Toates,
F.M. 1974. Computer
tional control U&off,
system
P.E. 1988.
ID5:
simulation
analysis
and the homeostatic
(McFarland,
an incremental
ID3.
control
of behaviour.
In Motiva-
ed.). Academic Press. Proc. 5th Int. Conf.
on Machine
learning.
Morgan
Kaufmann. Westman,
R.S. 1977.
Quantitative
Environmental
Methods
languages and the functional
in the Study of Animal
Behavior
(Hazlett
bases of animal ed.). Academic
behavior. Press.
In