behavior models
DESCRIPTION
TRANSCRIPT
Instructor: David A. Townsend
Email:
Class Web Page:
RESCORLA-WAGNER THEORY: BACKGROUNDResearch by Rescorla and Kamin requires a
change in our conception of conditioning. Animals do not evaluate CS-US pairings in
isolation.Evaluation occurs against a background that
includes: unpaired presentations of CS and US, or presentations of other CSs, or general experimental context.
Conditioning is most likely to occur when evaluation of entire situation reveals that a CS is best available predictor of US.
RESCORLA-WAGNER THEORY: BACKGROUNDThis view of Pavlovian conditioning makes
conditioning process appear much more complex than it had seemed.
How do animals do it? By what means do animals keep track of CSs
and USs, estimate probabilities, compute probability differences, and make CRs to CS?
If such calculations do occur, then animals are likely to make them automatically.
We need an account of the mechanism by which such automatic calculation might occur.
RESCORLA-WAGNER THEORY: BACKGROUNDIn last 30 years, many different theories have
been advanced to explain rich details of Pavlovian conditioning.
Most influential account was that of Robert A. Rescorla and Allan R. Wagner in 1972.
All later theories have been responses to shortcomings of Rescorla-Wagner account.
So, we will focus on Rescorla-Wagner theory.
RESCORLA-WAGNER THEORY: PRELUDEKamin’s blocking effect set stage for Rescorla-
Wagner model.Blocking effect suggested to Kamin that USs
were only effective when they were SURPRISING or unpredicted by CSs.
However, USs were not effective when they were unsurprising or predicted by CSs.
Added CS was not associated with any change in US.
RESCORLA-WAGNER THEORY: OVERVIEWRescorla-Wagner theory explains complex
contingency analysis in terms of simple associations of the sort Pavlov envisioned.
It can account for most standard conditioning phenomena in Chapter 3 as well as many newer phenomena in Chapter 4.
Theory is also precise; specified in such clear detail that one can derive predictions about behavior in untested experimental situations.
Mathematically
RESCORLA-WAGNER THEORY: TUTORIALSIf you need help, then please consult:http://psy.uq.oz.au/~landcp/PY269/r-wmod
el/index.htmlhttp://www.biols.susx.ac.uk/home/Martin_
Yeomans/Learning/Lecture6.htmlhttp://www.psych.ualberta.ca/~msnyder/A
cademic/Psych_281/C5/Ch5page.html
RESCORLA-WAGNER THEORY: ACQUISITION
Standard conditioning curves are negatively accelerated.
Changes in conditioning strength are very substantial early in training.
But, as training proceeds, a leveling-off point, or asymptote, is approached.
Generally speaking, changes in strength of conditioning get smaller with each trial.
RESCORLA-WAGNER THEORY: ACQUISITIONNegatively accelerated learning curve suggests
that organism does not profit equally from each training trial.
How much one profits depends on how much one already knows: When one knows nothing, profits are substantial (US
surprise is high).When one knows a great deal, profits from further trials
are small (US surprise is low).
RESCORLA-WAGNER THEORY: ACQUISITIONRather than learning a fixed amount with each
trial, one learns a fixed proportion of difference between one’s present level of learning and maximum possible.
As difference gets smaller (as one learns more), amount of new learning produced by further trials gets smaller.
Rescorla-Wagner model simply builds in a mathematical expression that conforms to negatively accelerated learning function.
RESCORLA-WAGNER THEORY: ACQUISITIONHere is the equation:
∆Vn = K(— Vn-1)V, associative strength, is measure of learning.
It is a theoretical quantity.It is not equivalent to magnitude or probability of
any particular CR.But, it is assumed to be closely related to such
measures of conditioned responding.
RESCORLA-WAGNER THEORY: ACQUISITION∆Vn = K(— Vn-1)
K reflects salience of CS.K can vary between 0 and 1 (0 ≤ K ≤ 1).Bigger K, bigger change in V on any given trial.Thus, salient stimuli mean large Ks, which mean
large ∆Vs, which mean large changes in association from trial to trial.
Intensity
Sensory modality
Organism
Types of US’s employed ( belongingness)
RESCORLA-WAGNER THEORY: ACQUISITION∆Vn = K(— Vn-1) indicates that different USs support different
maximum levels of conditioning. Asymptote of conditioning will vary with US;
different asymptotes are reflected by different s. More intense US, higher asymptote of
conditioning, and higher . is always equal to or greater than 0 ( ≥ 0).
RESCORLA-WAGNER THEORY: ACQUISITION∆Vn = K(— Vn-1)Change in strength on Trial n (∆Vn) is
proportional to difference between and prior associative strength Vn-1.
Because V grows from trial to trial, quantity ( - Vn-1) gets smaller and smaller, so ∆Vn also gets smaller and smaller, generating a negatively accelerated learning curve.
Eventually, V will equal , so that ( - Vn-1) will be 0, and conditioning will be complete
(asymptote will be reached).
∆Vn = K(— Vn-1)∆ (change in) DeltaV, associative strength, is measure of learningK reflects salience of CS. ( Lambda) indicates that different USs support
different maximum levels of conditioning.Change in strength on Trial n (∆Vn) is
proportional to difference between and Vn-1 prior associative strength.
Acquisition Trials
∆Vn= light (CS)=0 k = .20 = associated
strength of shock = 100
First conditioning trial:
Light (CS) is paired with shock (US)
First Trial∆Vn=k( -Vn-1)∆Vn= .20(100-0)∆Vn = .20 units
∆Vn = K(— Vn-1)∆ (change in) DeltaV, associative strength, is measure of learningK reflects salience of CS and US. ( Lambda) indicates that different USs support different maximum levels of conditioning.Change in strength on Trial n (∆Vn) is proportional to difference between and Vn-1 prior associative strength.
Acquisition Trials
∆Vtotal= light (CS)=0
k = .20 = associated
strength of shock = 100
First conditioning trial:
Light (CS) is paired with shock (US)
Second Trial∆Vn=k( -Vn-1)∆Vn= .20(100-20)∆Vn = .16 units
∆Vn = K(— Vn-1)∆ (change in) DeltaV, associative strength, is measure of learningK reflects salience of CS and US. ( Lambda) indicates that different USs support different maximum levels of conditioning.Change in strength on Trial n (∆Vn) is proportional to difference between and Vn-1 prior associative strength.
Acquisition Trials
∆Vtotal= light (CS)=0
k = .20 = associated
strength of shock = 100
First conditioning trial:
Light (CS) is paired with shock (US)
Third Trial∆Vn=k( -Vn-1)∆Vn= .20(100-36)∆Vn = .12.8 units
∆Vn = K(— Vn-1)∆ (change in) DeltaV, associative strength, is measure of learningK reflects salience of CS and US. ( Lambda) indicates that different USs support different maximum levels of conditioning.Change in strength on Trial n (∆Vn) is proportional to difference between and Vn-1 prior associative strength.
Acquisition Trials
∆Vtotal= light (CS)=0
k = .20 = associated
strength of shock = 100
First conditioning trial:
Light (CS) is paired with shock (US)
Forth Trial∆Vn=k( -Vn-1)∆Vn= .20(100-48.8)∆Vn = .10.2 units
∆Vn = K(— Vn-1)∆ (change in) DeltaV, associative strength, is measure of learningK reflects salience of CS and US. ( Lambda) indicates that different USs support different maximum levels of conditioning.Change in strength on Trial n (∆Vn) is proportional to difference between and Vn-1 prior associative strength.
N= trial4, N-1= trial 3
Rescorla-Wagner Model
Calculations from the Rescorla-Wagner model show a mathematical relationship to the process of conditioning 0
10
20
30
40
50
60
trial0
trial2
trail4
Vtotal
Rescorla-Wagner Theory (1972)
Organisms only learn when events violate their expectations (like Kamin’s surprise hypothesis)
Expectations are built up when ‘significant’ events follow a stimulus complex
These expectations are only modified when consequent events disagree with the composite expectation
Surprise
First Conditioning Trial
Trial K ( - Vn-1 ) = ∆Vn
1 .5 * 100 - 0 = 50
0
50
0
20
40
60
80
100
0 1 2 3 4 5 6 7 8
Trials
Ass
oci
ati
ve S
trength
(V
)
∆Vn = K(— Vn-1)∆ (change in) DeltaV, associative strength, is measure of learningK reflects salience of CS and US. ( Lambda) indicates that different USs support different maximum levels of conditioning.Change in strength on Trial n (∆Vn) is proportional to difference between and Vn-1 prior associative strength.
Second Conditioning Trial
Trial K ( - Vn-1 ) = ∆Vn
2 .5 * 100 -50 = 25
0
50
75
0
20
40
60
80
100
0 1 2 3 4 5 6 7 8
Trials
Ass
oci
ati
ve S
tren
gth
(V
)
Third Conditioning Trial
Trial K ( - Vn-1 ) = ∆Vn
3 .5 * 100 -75 = 12.5
0
50
75
87.5
0
20
40
60
80
100
0 1 2 3 4 5 6 7 8
Trials
Ass
oci
ati
ve S
tren
gth
(V
)
4th Conditioning Trial
Trial K ( - Vn-1 ) = ∆Vn
Trial c (Vmax - Vall) = ∆Vcs
4 .5 * 100 - 87.5 = 6.25
0
50
75
87.593.75
0
20
40
60
80
100
0 1 2 3 4 5 6 7 8
Trials
Ass
oci
ati
ve S
trength
(V
)
Vall
∆Vcs = c (Vmax – Vall)
5th Conditioning Trial
Trial K ( - Vn-1 ) = ∆Vn
5 .5 * 10 - 93.75 = 3.125
0
50
75
87.593.7596.88
0
20
40
60
80
100
0 1 2 3 4 5 6 7 8
Trials
Ass
oci
ati
ve S
trength
(V
)
6th Conditioning Trial
Trial K ( - Vn-1 ) = ∆Vn
6 .5 * 100 - 96.88 = 1.56
0
50
75
87.593.7596.8898.44
0
20
40
60
80
100
0 1 2 3 4 5 6 7 8
Trials
Ass
oci
ati
ve S
trength
(V
)
7th Conditioning Trial
Trial K ( - Vn-1 ) = ∆Vn
7 .5 * 100 - 98.44 = .78
0
50
75
87.593.7596.8898.4499.22
0
20
40
60
80
100
0 1 2 3 4 5 6 7 8
Trials
Ass
oci
ati
ve S
trength
(V
)
∆Vn = K(— Vn-1)∆ (change in) DeltaV, associative strength, is measure of learningK reflects salience of CS and US. ( Lambda) indicates that different USs support different maximum levels of conditioning.Change in strength on Trial n (∆Vn) is proportional to difference between and Vn-1 prior associative strength.
8th Conditioning Trial
Trial K ( - Vn-1 ) = ∆Vn
8 .5 * 1 - 99.22 = .39
0
50
75
87.593.7596.8898.4499.2299.61
0
20
40
60
80
100
0 1 2 3 4 5 6 7 8
Trials
Ass
oci
ati
ve S
trength
(V
)
1st Extinction Trial
Trial K ( - Vn-1 ) = ∆Vn
1 .5 * 0 -99.61 = -49.8
Acquisition
0
20
40
60
80
100
0 1 2 3 4 5 6 7 8
Trials
Ass
oci
ati
ve S
trength
(V
)
Vall
Extinction
99.61
49.8
0
20
40
60
80
100
0 1 2 3 4 5 6
Trials
Ass
oci
ati
ve S
tren
gth
(V
)
2nd Extinction Trial
Trial K ( - Vn-1 ) = ∆Vn
2 .5 * 0 -49.8 = -24.9
0
50
75
87.593.75 96.88 98.44 99.22 99.61
0
20
40
60
80
100
0 1 2 3 4 5 6 7 8
Trials
Asso
ciativ
e St
reng
th (V
)
Vall
Acquisition
0
20
40
60
80
100
0 1 2 3 4 5 6 7 8
Trials
Asso
ciativ
e St
reng
th (V
)
Extinction
99.61
49.8
24.9
0
20
40
60
80
100
0 1 2 3 4 5 6
Trials
Ass
oci
ati
ve S
tren
gth
(V
)
Extinction Trials
Trial K ( - Vn-1 ) = ∆Vn
3 .5 * 0 - 12.45 = -12.46
4 .5 * 0 - 6.23 = -6.23
5 .5 * 0 - 3.11 = -3.11
6 .5 * 0 - 1.56 = -1.56
Less and Less surprising
Hypothetical Acquisition & Extinction Curves with K=.5 and = 100
Extinction
99.61
49.8
24.9
12.456.23 3.11 1.560
20
40
60
80
100
0 1 2 3 4 5 6
Trials
Associ
ati
ve S
trength
(V
)
Acquisition
0
20
40
60
80
100
0 1 2 3 4 5 6 7 8
Trials
Asso
ciativ
e St
reng
th (V
)
Acquisition & Extinction Curves with c=.5 vs. c=.2 ( = 100)
Acquisition
0
20
40
60
80
100
120
0 1 2 3 4 5 6 7 8
Trials
Ass
oci
ati
ve S
trength
(V
)
c=.5c=.2
Extinction
0
20
40
60
80
100
120
0 1 2 3 4 5 6
Trials
Ass
oci
ati
ve S
trength
(V
)
c=.5c=.2
RESCORLA-WAGNER THEORY: COMPETITIONKey feature of Rescorla-Wagner model is how it
explains conditioning with compound stimuli comprising two or more elements.
Associative strength of a compound stimulus is assumed to equal the sum of associative strengths of elements.
VAX = VA + VX
Here, A and X may have different saliences, K and M, respectively.
∆Vn = K(— Vn-1)∆ (change in) Delta V, associative strength, is measure of learning K reflects salience of CS. ( Lambda) indicates that different USs support different maximum levels of
conditioning. Change in strength on Trial n (∆Vn) is proportional to difference
between and Vn-1 prior associative strength.
A,X symbols used for multiple CS’sSalience with multiple CS’s: A=K,
X=M
RESCORLA-WAGNER THEORY: OVERSHADOWINGVAX = VA + VX
How can theory account for overshadowing?With equally salient stimuli, VX would attain
only .50 rather than 1.00 if X alone were trained--mutual overshadowing.
Increases in salience of A would further reduce VX from .50 toward .00.
If salience of A (K) is very high and salience of X (M) is very low, then overshadowing should be complete.
Eyeblink Conditioning: OVERSHADOWING
Training: Tone/light + ShockTone = Eyeblink CRLight = ?
No CR to light
Corneal Air PuffElicits Eyeblink Response
Corneal Air PuffGiven with Tone
Tone Given AloneElicits Eyeblink Response
Overshadowing:
Overshadowing Whenever there are multiple stimuli or a compound stimulus, then ∆Vn = Vcs1
(K) + Vcs2 (M)
Trial 1:∆Vnoise = .2 (100 – 0) = (.2)(100) = 20∆Vlight = .3 (100 – 0) = (.3)(100) = 30Total ∆Vn = ∆ (K)Vnoise + ∆Vlight = 0 +20 +30 =50
Trial 2:∆Vnoise = .2 (100 – 50) = (.2)(50) = 10∆Vlight = .3 (100 – 50) = (.3)(50) = 15Total ∆Vn = Vn-1 + ∆Vnoise + ∆Vlight = 50+10+15=75
∆Vn = K(— Vn-1)
Noise= 30
Light= 45
Overshadowing
Trial 1: DVA = .40(100 – 0) = 40
DVx = .10(100 – 0) = 10
Trial 2: DVA = .40(100 – 50) = 20
DVx = .10(100 – 50) = 5
T2:
A=60
X=15
RESCORLA-WAGNER THEORY: BLOCKINGVAX = VA + VX
How would theory account for blocking?With equally salient stimuli, VX would be only .50
rather than 1.00 if AX only were trained.Prior training with A would further reduce VX,
because VA would already be substantial before AX trials were introduced.
Extensive training with A should lead to complete blocking of X.
Blocking
Group Phase 1 Phase 2 Phase 3
ExperimentalGroup (blocking)
ControlGroup
A US
Nothing
AB US
AB US
Test B
Test B
Same # trialsContiguityContingency
RESCORLA-WAGNER THEORY: BLOCKING
Acquisition
0
20
40
60
80
100
0 1 2 3 4 5 6 7 8
Trials
Ass
oci
ati
ve S
trength
(V
)
A
X
?
The Rescorla-Wagner associative model of conditioning is based upon four assumptions that refer to the process by which the CS and UC gain
associative strength (1) a particular US can only support a specific
level of conditioning, (2) associative strength increases with each
reinforced trial, but depends upon prior conditioning,
(3) particular CSs and US can support different rates of conditioning and
(4) when two or more stimuli are paired with the UC, the stimuli compete for the associative strength available for conditioning.
RESCORLA-WAGNER THEORY: CONTINGENCYTheory can also explain animals’ ability to detect
different degrees of contingency between CS and US.
Recall that fear of a CS for shock is a direct function of contingency between CS and US.
When contingency between events is zero, no learning of fear to CS occurs.
But, does a rat really compute probabilities to form a judgment of contingency?
Not according to Rescorla-Wagner model.
RESCORLA-WAGNER THEORY: CONTINGENCYTo explain contingency
sensitivity, Rescorla-Wagner theory makes use of background or contextual stimuli as Pavlovian predictors.
(Context = A discrete CS)Such contextual stimuli
themselves can compete with CSs for association with USs.
Case of random presentations of CS and US provides a useful illustration.
TRAINING
CONTEXTUALTEST
CUED TEST
Cue-plus-contextual Fear Conditioning
A.
TRAINING CONTEXTUALTEST
Context Alone Conditioning
Context Discrimination
TRAINING
TEST
Context 1
Context 2
Context 1
B.
C.
Variations of Fear Conditioning
RESCORLA-WAGNER THEORY: CONTINGENCY Random training can be seen to represent blocking with two kinds of
trials: A (context)-US [relatively frequent] AX (context plus CS)-US [relatively infrequent]
As animal receives frequent A-US pairings, VA (and hence VAX) approaches asymptote.
As VAX approaches asymptote from frequent A-US pairings, VX can receive no further increments and little responding to X will be observed despite occasional AX-US pairings.
US
CS
unpaired
time0.5 s
RESCORLA-WAGNER THEORY: CONTINGENCYSo, blocking is basic to effect of random CS and
US presentations. Contextual cues are present whenever US
occurs in absence of CS; contextual cues thus acquire excitatory strength.
On trials when CS is paired with US by chance, contextual cues are present as well.
So, context replaces Stimulus A in blocking example and randomly presented CS replaces Stimulus X.
RESCORLA-WAGNER THEORY: INHIBITIONIn Chapter 4, we saw that conditioning can be
either excitatory or inhibitory. At first glance, it is not obvious that Rescorla-
Wagner theory can explain inhibition. Inhibition requires a V that is less than zero; but,
none of the variables in the equation can ever be less than zero.
How can V become negative when none of the terms contributing to V
can be negative?
Time to Leave:
Zero
X Axis
Y A
xis
0
1