Flashcards in Rescorla Wagner and Blocking - tutorial 1 Deck (20)
the basic equation
dV = ab(l - SV)
So learning (changes in associative
strength, V) depends on two constants (a, b) and the prediction error (l - SV).
learning rate parameter determined by the salience of CS;
learning rate parameter determined by the salience of the outcome.
asymptote of associative strength (V).
l has a positive value on episodes when the outcome is US with its magnitude being determined by the effectiveness of US.
sum of the associative strengths of all CSs present on the learning episode.
often referred to as the prediction error or D term because it reflects the extent to which the combined associative strength of all CSs present on the learning episode (SV) predicts the outcome (l) on that episode.
When (l - SV) is zero, the outcome is fully predicted and no further changes in associative strength (dV), i.e., no learning, occurs.
As l is zero when no US is presented, non-reinforcement yields a negative D term and hence a decrement in V, thus producing extinction.
Say you start with one pairing of a CS, followed by a US.
If ab = .5, and the US supports an asymptotic associative strength of 1, and at the start V(there’s only one CS in this example)=0.
dV = ab(l - SV) = .5x(1-0) = .5
So the next V will be 0 + 0.5 = 0.5, up from 0. Now let’s have
another trial where CS and US are paired.
dV = ab(l - SV) = .5x(1-.5) = .25
So the next V will be 0.5 + 0.25 = 0.75, up from 0.5.
Bear in mind that associative strength as computed by Rescorla-Wagner does not necessarily translate directly into responding – but we usually assume that the relationship between the two is monotonic.
Blocking (Kamin, 1969)
Stage 1: CS noise + shock --> CR (fear)
Stage 2: CS noise + light + shock --> CR (fear)
Test: light alone --> little or no (fear) conditioned suppression (0.45)
Suppression ratio =
CSresponding/CSresponding + PreCSresponding
Control condition (no stage 1) gives suppression of 0.05 to the light on test.
RW explanation of blocking
We’ll start with a CS who’s V is already 0.75, the one we trained earlier! That’s the noise then.
We’ll add another, of equal salience (so a is the same), that’ll be the light. It’s novel, so its V is 0.
dV (either CS) = ab(l - SV) = .5x(1-.75) = .125
This is because SV is the same for both of them.
It’s .75 + 0 = .75.
V(noise) is now .875 =(.75+.125).
V(light) is now .125 = (0 + .125).
On the next trial SV is 1, so conditioning stops. Weak associative strength for the light.
But say instead that both noise and light were novel, and started from 0 (control condition).
On the first trial:
dV (noise or light) = ab(l - SV) = .5x(1-0) = .5
And on the next trial SV = 1, so conditioning stops (I’ve chosen my values carefully here to make the maths easy).
The light ends up with an associative strength of 0.5, which is much
bigger than 0.125 – and that’s blocking!
RW explanation of overshadowing
Conditioning noise and light together gives them both a V of 0.5.
But conditioning them separately will give them a V of 1 (why?).
Hence, overshadowing due to conditioning in a compound.
The objection often made to demonstrations of this type is that people know what they’re doing, and you are testing their memory and cognitive inference rather than their learning.
Maybe they just work it out?
Participants may have reasoned something like this:
- A+=> A is effective.
- So AB+ => B may not be effective as it could all be A
- Give A more credit, as it’s in both.
CD+=> means that C and D could both be responsible. Split the credit between them.
Hence A rating higher, B rating higher lower, and C and D ratings in between.
E+ => it’s all E. Give high rating.
Or it could be associative learning using something like RW
Beckers et al.
Their big idea is to assume that rats reason, in the same way that undergraduates might have done while doing the allergy blocking experiment, and then to try and influence the assumptions that underpin reasoning in rats.
This states that if cue A signals an outcome (say food) and cue B signals the same outcome, then A and B presented together should signal a stronger outcome (more food in this case).
Their analysis predicts that this assumption is needed to explain blocking in rats and humans.
In the earlier experiments with humans this assumption was implicit in the inference that if A was effective then AB+ => that B was ineffective.
If B had been effective then more of an outcome should have occurred.
additivity and blocking
If an animal is exposed to the sequence of trials A+|AX+ then if it expects effects to add, it can deduce that X is ineffective - and blocking will be exhibited
If it doesn’t expect additivity, then X might be an effective cue and blocking should be reduced
The idea is that the pretraining with C and D in the experimental group makes the animal expect nonadditivity, and so it should not show blocking (but will do so in the Control)
Their results supported this prediction.
The animals trained not to expect additivity (subadditive) showed no blocking
The idea is that the exposure to shock allows the
animal to determine if an outcome is maximal
(can’t increase) or not.
If it can’t increase - then
can’t expect additivity and so shouldn’t get blocking
exp 3 results
Their results supported this prediction.
The animals trained not to expect additivity (maximal) showed no blocking