Behavioural Elsevier
Processes,
CLASSICAL Nestor ment
14 (1987) 277-289
CONDITIONING,
A.
SIGNAL
Schmajuk,
Center for 111 Cummington
of Mathematics,
(Accepted
277
10 February
DETECTION,
AND EVOLUTION
Adaptive Systems, Boston University, DepartStreet, Boston, Massachusetts 02215 USA
1987)
ABSTRACT N.A. Schmajuk. Behav.
Classical
Processes
conditioning,
14:
signal
detection,
and evolution.
277-289.
Strength of classical conditioning is increased either by increasing discriminability of the conditioned stimulus (CS) from the background, or by increasing contingency between conditioned and unconditioned stimuli (US). Classical conditioning can be regarded as a decision process in which the subject has to decide whether or not to respond with a conditioned response in the presence or absence of the CS. According to modern evolutionary theories, it might be assumed that this decision process maximizes the trade-off between costs and benefits. By assuming that the decision rule maximizes expected benefit, the empiric,al relationship between contingency and the strength of classical conditioning is theoretically derived. In addition, when the decision rule is incorporated to a signal detection paradigm, theoretical results describing the relationship between C’S discriminability and CS - US contingency
with the strength
of classical
conditioning
are in agreement
with experimental
data.
Introduction Discriminability gency
bet,ween
of the major Strength
conditioned
variables
and, except
Conditioning rather
than
with
0376-6357/87/$03.50
as a function
was a function
simple
temporal
of the noise
CS
independent,
of CS CS
- ITS ront,iguity.
as some
discriminability
intensity.
from
Using
either
(1965)
found
from the background
noise.
Kamin
of the change in background from the absolute
with
Q 1987 Elsevier Science Publishers
have been identified
as a CS,
discriminability
also to increase CS
intensity
and contin-
conditioning.
absolute
of the magnitude
in intensity,
seems
stimuli
to increase
as a function
strength
(US)
from the background,
of classical
by increasing
for small changes
(CS)
the strength
in background
increased
of conditioning
seems
than
or an increase
that, conditioning St,rength
rather
stimulus
and unconditioned
determining
of conditioning
t,he background. a decrease
of the conditioned
inrreasing Rcsrorla
CS
noise
intensity.
- 1i.S contingency,
(1068)
B.V. (Biomedical
CS
suggested
Division)
that
278
conditioning
is attained
when the CS
is a reliable
predictor
of the IJS,
receiving
the US in the presence
of the CS
(P(USlCS))
is greater
receiving
the US in the absence
of the CS
(P(USlCS)).
R escorla
conditioning
is obtained
conditioning
from negative
from
non-contingency
(1974)
proposed
CS
and US,
CS.
Gibbon
from positive
contingencies
(P(USlCS)
et al. (1974)
that
detect
learning
terms
of costs
Smith,
1978)
about
mechanisms,
propose
that
excitatory inhibitory
and no conditioning
Berryman,
a statistical
of the reliability
and Thompson
correlation
of the prediction of conditioning
natural
between
of the i/S
by the
was proportional
theories
to
rules can be theoretically In addition,
the probability
and Davies, maximally
trade-off
animals
increasing
values
analyzed
efficient
in
Maynard behavioral
costs and benefits
theories, of CS
to
The survival
1981;
between
to optimization
the present
study
derived by assuming
when the decision operating
of emitting
1972; Testa,
to enable
the strategy
of
discriminability
and
for survival.
predict,
benefit.
and Kalat,
can be quantitatively (Krebs
has chosen
According
value with
is optimal
receiver
behaviors, theories
selection
Rozin
evolution
in their environment.
governed by a particular
tingency
paradigm,
other
net benefit.
- US associative
As optimization
1971;
through
causal relationships
among
i.e., those strategies
CS
1980; Revusky,
Optimization
that
CS - US contingency
tection
< P(USlCS)),
contingency,
have been shaped
and benefits.
gives the maximum
increasing
argued
> P(USlCS)),
Gibbon,
that the strength
(Dickinson,
mechanisms
and store information
strategies;
(P(US(CS)
of of
of the contingency.
value of learning
that
measure
suggested
It has been suggested 1974)
(P(USlCS)
= P(US(CS)).
use of the root mean square
aa a quantitative
the magnitude
contingencies
t,he probability
than the probability
shows that
a CR under different
(1968)
that these rules maximize
rule maximizing
characteristic
Rescorla’s
(ROC)
benefit
is applied
con-
expected
to a signal
de-
curves can be used to determine
CS - US contingencies
and CS
discrim-
inabilities.
Contingency
and derision
Classical the animal
conditioning decides
theory
can be regarded
whether
the presence
of the CS superimposed
noise alone.
This decision
CS
is present
by the animal, world,
the US
or absent
process
(C’s)
or absent
detection
with a conditioned
to background is characterized
or not emmited
(US) (Egan,
problem
response
(CR)
noise or in the presence by two different
on the noisy background;
the CR is emitted is present
as a fundamental
or not to respond
(cri);
either
responses
and two different and Swets,
in
of background
values of evidence,
two alternative
1975; Green
in which
states
1966).
the
selected of the
279
In t.hcLfundamental hit, (the animal a CR
with
detection
but
no
is delivered),
US
t,he IJ.5’ is delivered),
and correct
the TJS is not delivered). different
pay-off
Table
values should
1 shows
value:
r/(US,
the CR),
conditioning
that
pay-off
(1982)
deal with
the forthcoming
rival male.
and l/S,
lJS
the US
V(US,CR)
with
Hollis
the prrsent,ation
Hollis’( 1982)
the CS
the lJS
proposed
plus the cost of
(the cost of eliciting
interactions
Hollis (1984)
a
when the CS
view is applied,
on the basis the animal
with predators,
enjoy an adaptive
precedes
it is V(US,CR)
< V(IIS,CR); be less than
alone.
Payoff
matrix
for a classical
I conditioning
experiment
DE
_-~_y(US> CR) _-
CR, minus the cost of receiving
the benefit
a CR whenever
If
V(US,CR)
sign) of producing
of Equation
rule:
> V(U_S, CR)
to the CS.
large number
the CR,
rule to maximize
+ P(USICS)V(US,CR)
in the decision
P(USlC.5) ~_ P(USlCS)
is to produce
produces
A decision
+ P(~~f$‘S)V(US,CR)
P(USlCS)V(US,CR)
results
the animal
the CR.
can be rewritten
P(USICS)V(US,CR)
>
when
does not produce
when the CS
CR) (Egan,
E(CS,
matrix:
when the animal
the CR.
than
Increasing
of t,he t,he
required
for responding.
of P(IrSICR)
required
for
When
the
responding. So far, a decision CS
is absent,
when
rule has been derived for the case when the CS
two expected
the animal
produces
the CR.
A decision
produce
CR whenever
values can be computed the CR,
and E(CRICS),
rule to maximize
E(CS,(,‘R)
the expected
terms
results
in the decision
P(USlCS) m~Lw P(USlCS)
E(CRICS).
does not produce
when the CS
is absent
is t,o
as
+ P(USICS)V(US,CR)
> P(USICS)V(TJS.CR) Reorganizing
when the animal benefit
This can be written
> E(CS,CR).
r(r’slcs)V(r!s,CR)
is present.
upon the pay-off matrix:
’
+ P(UslCS)V(US,CR) rule:
[31
If
V(US,CR)
- V(TJS, CR)
V(US,CR)
- V(TJS, CR) ’
PI
281
respond
with a C’R during
The animal
should
the absence
not respond
of the CS.
in the absence
of the C’S when
P(US(CS) < V(US,CR) - V(USYCR) _-__ P(US(CS) C’onceptually,
the decision
the rat,io between absence
of receiving the ratio
and those of not receiving
with
the CR if
the US
in the
between
for not responding.
Excitatory CS
than
as: Do not respond
the cost and the benefit of the CR. Increasing the benefit produced by the CR decreases the value of P(USlCS) required for not responding. Increasing the cost of producing the CR increases the value of P(USlCS) required
is smaller
- V(U?’
rule can be expressed
the probabilities
of the CS
V(US,CR)
conditioning
is attained
when
but not in its absence.
Therefore,
the conditions
should
be simultaneously
satisfied.
Equations
P(US~CS)
P(USlCS)
7 is identical
Therefore
the criteria
the absence
of the CS
Inhibitory
for obtaining is equivalent
rondit,ioning
but not, in its presence. 2 and 5 should
to Rescorla’s
be simultaneously
P(IJSjCS)
9 is identical
= 1 - P(USlCS),
we have:
for effective
opposite
excitatory
benefit
for establishing
when the animal
conditions
171
net behavioral
to the criterion
responds
conditioning.
in the presence
classical
in the absence
to those represented
and
conditioning. of the CS
by Equations
satisfied:
= 1 - P(US(CS),
< P(USlC.9) ~., P(USlCS) --and P(USlCS)
P(USlCS) Equation
condition
the maximum
P(US(CS) ~_____ P(USlCS) Replacing
161
> P(USlCS).
(1968)
is attained
Therefore,
of the 2 and 5
P(U.qCS)
P(US(CS) Equation
in the presence by Equations
2 and 5 imply:
and P(US(CS)
= 1 - P(fiSlCS),
responds
represented
> P(USlCS) ~.__._.
__.._~
P(US(CS) Replacing
the animal
to Rescorla’s
(1968)
= 1 - P(USlcS),
we have:
I91
< P(USlCS). condition
for effective
inhibitory
conditioning.
282
Contingency,
discrirninebility,
In the preceding
section
baaed has only two discrete evidence
has a continuous
z constitutes
and signal detection
we have assumed values:
CS
on which the decision
Pyz)
Bayes’
P(xlUS)
the posterior
-P(zpJS)P(US)
Equation
12 gives the decision
likelihood
ratio,
P(xlUS)
If the right-hand written
probabilities
P(USls)
and P(-iislx),
we obtain
- V(US,CR) -~ - V(US,CR)
I111
P(US)V(tiS,CR) _~ ~.. .~~.
-~~_~~_ V(vS,CR) ~~.~ -. CR) - V(US, CR)
> ._~
rule for responding
= P(xIUS)/P(xlUS), and P(xlUS)
(121
based on the evidence
of the observation.
is discussed
term of Equation
z, in terms of the
The meaning
of the prior
the decision
rule can be
below.
12 is designated
L(zo),
as
Equation
13 divides the L(x)
each x that belongs the second
> L(x0).
[I31
axis into two intervals:
L(x)
to the first interval no CR is emitted,
< L(xo)
and L(z)
2 L(zo).
For
and for each z that belongs
to
interval a CR is emitted.
and signal-plus-noise
distributions
under
different
contingencies
We assume that the CS signal is added to noise with a normal intensity density function along a physical
continuum.
4 =
always
1, the CS
density
a stimulus function
of intensity
of intensity
with mean rncs
signal detection
in the presence
= 1. Under these conditions
P(CSlUS) receiving
In a standard
appears
two density functions
x when US is produced,
p(zlTf~),
u,.
procedure
of the 1JS;
x when US is produced, and variance
’
WI
V(US,CR)
L(x)
Noise
When
2 can be rewritt,en as
-. CR)
> V(US,CR) -
P(US)V(US,
probabilities
such as light or sound.
Equation
is
that the
terms: L(x)
L(z)
the decision
to assume
- V(US,CR)
and P(zlUS),
P(x~US)P(US)
and after rearranging
is based,
V(US, CR) -I/(US,
law for replacing
on which
It is also possible
> V(US,CR)
WJS I4 Applying
that the evidence
and CS.
value, z, along a physical dimension
the evidence
by the prior probabilities
theory
i.e.,
appear.
p(zll/S),
The probability
is given by a normal
with contingency
P(CSlUS)
= 1 and
The probability
of
is given by a normal of receiving
a stimulus
density function
with
283
mean mcs
and variance
physical intensity When
u,.
The distance between
is the discriminability
contingency
is different
is produced,
p(z]US),
combination
of two density functions,
in proportions
given by P(CS]US)
P(C.YS]US)~O&
of receiving
a stimulus of intensity
is given by a normal density function
has mean .st = P(CSIUS)mcs
z when US
that, results from the linear
one with mean rnCS and the other with mean rnz, respectively.
and P(CS]US),
+ P(CS(US)mcs,
For the reasons presented
The resulting function
and variance UT = P(CS]US)2a&s
= (2auf)):
exp[-(s
+
in the preceding
- sr)‘/2uf].
paragraph,
stimulus of intensity z when US 1s produced, p(zlUS), with mean .s2 = P(CSIUS)mcs
I141
the probability
of receiving
IS . g iven by a normal density function and variance ui = P(CSIi!???)2u&
+ P(CSIUS)mm,
a
+
Therefore,
P(CsIUS)2u$,s.
p(zlUS)
and p(zlUS)
p(zlUS)
Assuming
rnCS =
P(CSIUS))d’/2,
z)
P(CSjUS)
=
- s2)2/2ul].
I151
in the expression of the likelihood
=
ratio,
e3?!~rk2drL2~~ exp[-(z
d’/2 and rnc:g =
and s2 = (P(CSlU.5’)
For the case P(C5’lCJS)
exp[-(2
= (2~4~~
q
placing
in
(Hays, 1973, page 314). Therefore, p(zlUS)
Replacing
i.e., the difference
than 4 = 1, the CS may imply either the US or the
We will assume that the probability
US.
rn~s and rnz;
of the CS.
-d’/2,
mean values are sr =
(P(CSlUS)
-
- P(CSIUS))d’/Z.
= P(cSl~s),
1 - P(CSIIIS),
I161
- 42/2u;1
= (I#+ 1)/2 (see Appendix).
P(CSlUS) P(CSlUS)
=
1 - P(CS(US),
and (I,
=
Re1, we
obtain L(z) where
= exp[@‘d’z],
4’ = 2$/$2 + 1. 4’ is an increasingly
4 5 1. The effect of changing contingency
monotonic
4 is equivalent
function
of 4 in the range 0 5
to varying t,he effective
signal detection
paradigm.
disrriminability
d’. c$ smaller than 1 implies a decrease in the effective
Figure
4 equal to 1 implies that L(z)
1171
1 shows likelihood
4 = 1, the likelihood
ratio L(z)
ratio is L(s)
is determined
by the stimulus
d’.
as a function of 2 for different values of 4. When
= cd”, which is the value obtained
theory for two normal density functions
d’ in a
in signal detection
(Green and Swets, 1966, page 60). For 4 = 0, the
284
0
’
_d;2
Figure 1. Two normal -
density
functions,
‘0
d)2
one for the CS
(mean
d/2), and other
for the
CS (mean -d/2) are shown. When different CS - US contingencies are applied to these density functions, different likelihood ratios L(z) as a function of CS intensity (z) and contingency (4) are obtained. When L(r) equals L(zO) at zO, z, divides the 2 axis in
two intervals. The animal responds with a CR t,o z values in one interval and with CR to z values in the other interval. The shaded area under the P(CS) function represents P(CRICS). For this illustration Q = .2. Higher values of 4 imply lower values of zc,, and therefore greater values of P(CRIC.5’). likelihood
ratio
L(z)
for I$ with values
is one for every value of z. Intermediate
.8,.6,.4,
and .2. In all cases
values
of L(z)
is an exponentially
L(X)
are obtained
growing
function
of z. The CS CS,
response
density
P(CRlCS),
Figure
in the presence
for I > z,.
is the integral
1 shows
decreasing
rate
function
that
under
for a given
the areas under P(CS)
of the CS,
Correspondingly, the CS
L(z,),
z,
P(CR(CS), the response
density
increases
and the P(CS)
is the integral
under
rate in the absence
function
within
as contingency
and consequently
the
of the
the same
interval.
decreases,
thereby
decreasing
P(CRjCS)
and P(CRICS). Integrals normal
of normal
integral
functions
cannot
be evaluated
with mean zero and unit variance
in closed
can be obtained
form,
but
the st,andard
from tables.
Therefore,
285
P((.‘RICS)
and P(CRICS)
and determining Application
compared
Equation dat,a from
tioning during CS
the CS
onset.
is assumed between
to have
the baseline
Figure between
period
2 shows
theoretical
derived
(1968)
values
for cases
and Stein,
equals
L(s,),
ratios,
A/(A
as the difference
probability
one)
rate of responding
and
(1958)
ex-
have been
d’ was selected
was set equal to 20. Strength
equal
of condi-
+ B), where A is the rate of responding
between
during a comparable the baseline
P(CRICS).
period
prior to
rate of responding
(which
B is computed
as the
difference
and P(CRIcs).
and experimental
and experimental
= P(CSlUX),
and Brady
D’lscriminability
of P(CRICS).
and B is the rate of responding
theoretical
L(.z)
with P(CS1U.S)
Sidman,
to 2.5, and L(s,)
as suppression
A is computed
t at which
functions.
data
on was set equal
is expressed
by finding
the density
17 has been Rescorla
to t,he theoretical
to 10, variance
be obtained
under
to cxprrirnrnt~al
Although perimenhal
might
the integrals
values
values is significant,
as a function
of 4.
Correlation
rzY = .97, L(5) = 6.86,
p < 0.01.
theoretical experimental
.2
.4
.6
.8
1.
contingency Figure 2. Suppression rat,ios as a function of 6. Experimental values were obtained from Rrscorla (1968) (d = 0 to .6), and from and Stein. Sidman, and Brady’s (1958) study recalculated by Gibbon ct ~1. (1974) m . suppression rat,io form (4 = 1).
286
ROC
St/u&d Standard different
curves
ROC
values
of P(CRIUS) ext,reme
curves
of 4.
might
In signal
and P(CRIUS)
values.
When
curves
contingency
detection
between
L(z,).
2,
by entering
theory
obtained
q5 = 1, ROC
P(C.5’) curves for different P(CS)
be used to determine
when
curves
Since
and infinity,
P(CRICS)
a ROC
and
is the set, of different
IS varied continuously
L(r,)
provide
the area
ROC
curves
provide
ROC
curves
can
under
for
P(CRIC.7)
values
between
the P(CS)
the area under
its
and the and
P(CS)
be used for different
values
of
a R,OC curve with D = q5’d’.
Discussion In this subject
paper,
classical
condit,ioning
has to decide whether
or absence assumed By
or not to respond
of the conditioned that
this decision
assuming
that
corla’s
(1968)
values
of contingency
maximum
process
that
benefit
for establishing
have been
signal detection the model linear limited
to the
description
Therefore
theories,
it is
costas and benefits. expected
increases
the criteria
and the absence
and CS-IJS
experimental
lower range
in the presence
benefit. with
Res-
increasing
for obtaining
of the CS
the
is equivalent
conditioning.
derived
contingency
contingency
by applying
on the strength
the optimal
data.
As Gibbon
values.
et al.
(1974)
experimental
The
present
of classical
decision
I erent contingencies f or d’ff
and Rescorla’s
of contingency
also provides
of conditioning,
a good description
as described
rules to a
computed
with
pointed
out, a
suppression model
in CS discriminability
intensities
of t,he (,‘,S distribution,
in t,hr presence Several
ratios
provides
is
a good
suggested
between
paradigms
and Yarensky,
of the effect of CS (1965).
discriminability
A.5 in a signal detection
thereby
increasing
on the paradigm,
ratio for the lower
the t,ot,al probability
of responding
of the CS.
authors
int,eractions
by Kamin
(d ’) increase the value of the likelihood
increments
tioning
of conditioning
the
for all values of contingency.
The model strength
between
response
in which
evolutionary
maximize
strength
Values of P(CRICS)
tit. well Rescorla’s
relationship
process
between
derived.
classical
theoretically
paradigm.
strategies
in the presence
The effects of CS discriminability conditioning
to modern
the trade-off
condit,ioning
postulate
as a decision
with a conditioned
According
maximizes
is theoretically
net behavioral
to the criterion
stimulus.
classical
empirical
is regarded
that
signal
rate of reinforcement (Davison
1977).
Nevin
and Tusting, (1981)
detection
theory
can be applied
and stimulus
discriminability
1978:
1981;
Nevin.
Nevin,
to describe
the
in opf’rant, rondiJenkins,
b,h owed t,hat,, whrn changc>s in disrriminability
Whitt,akrr. arr rqj-
287
resented
as changes
likelihood
ratio
in d’, and changes
inabilit,y
and contingency
ing case,
the present
theory
in contzingency
R.OC-l’k1 e curves describe
L(r),
in operant
paper
conditioning
between
As in the operant
ROC curves
CR probability
as changes
interactions
paradigms.
shows that standard
can be used for describing
are represented
adequately
in the discrim-
rondition-
derived from signal detection
as a function
of discriminability
and
contingency. As ment,ioned maximally
above,
efficient
optimization
behavioral
theory
t,o operant
derived
from optimal
conditioning foraging
In the case of classical has chosen to the
behavioral
CS.
This
discriminabilit,y the principles
rules.
The version
value,
that
optimization
is confirmed
is an account
an explanation
law can
that natural
be
that
selection
when to respond the effects
of C’S
can be derived
(1983)
of the rules used in classical
for vrhy animals not,ed,
(1983)
has pointed
was supported
in deciding
showing
Shettlcworth
is grateful
predict
efficient
conditioning
a description
author
has chosen optimization
matching
theories
on classical
learn,
of this manuscript,.
select,ion applied
from
t,heory.
approach
animals
(1970)
by our results
contingency
and Hinson
hew
natural
successfully
Herrnstein’s
of hour animals
luhy and
that
(1980)
that, are maximally
derision
optimality
.4s Staddon
Furthermore, hrtwcrn
conditioning,
of optimal
of survival
description
showing
strategies
prediction
assume
Staddon
models.
and C’S - US
The present in terms
theories
strategies.
learn
opt,imality
according
approaches
of the mechanisms out that
there
conditioning to particular
do not
involved
is no simple
provide
a
in learning. relationship
learn.
to Dr.
Dan
Bullock
Also to Cynthia
for his vahiable
Suchta
for typing
comments
on an early
the manuscript.
This paper
in part by NSF grant, IST-8417756.
REFERENCES Dickinson, A., 1980. Ilniversity Press. Egan.
.J.P., Press.
1975.
Condemporary
Signal
det,ection
animal theory