Keyboard Shortcuts?f

×
  • Next step
  • Previous step
  • Skip this slide
  • Previous slide
  • mShow slide thumbnails
  • nShow notes
  • hShow handout latex source
  • NShow talk notes latex source

Click here and press the right key for the next slide.

(This may not work on mobile or ipad. You can try using chrome or firefox, but even that may fail. Sorry.)

also ...

Press the left key to go backwards (or swipe right)

Press n to toggle whether notes are shown (or add '?notes' to the url before the #)

Press m or double tap to slide thumbnails (menu)

Press ? at any time to show the keyboard shortcuts

 

Dual Process Theory Opposes Decision Theory?

habitual process

Action occurs in the presence of Stimulus.

Agent is rewarded [/punished]

Stimulus-Action Link is strengthened [/weakened] due to reward [/punishment]

Given Stimulus, will Action occur? It depends on the strength of the Stimulus-Action Link.

‘goal-directed’ process

Action leads to Outcome.
 

Belief in Action-Outcome link is strengthened.

Agent has a Desire for the Outcome
 

Will Action occur? It depends on the Belief in the Action-Outcome Link and Agent’s Desire.

What are these beliefs and desires? How exactly do they lead to actions? I suggested that we can appeal to decision theory for an answer to both questions.
I suggested that we take game theory as elucidating the goal-directed process.

This book has ‘a philosophical end: elucidation of the notions of subjective probability and subjective desirability or utility’

(Jeffrey, 1983, p. xi)

key assumption:

Agents’ actions maximise their expected utilities.

If we understand this as instrumental action, we run into a problem.

This book has ‘a philosophical end: elucidation of the notions of subjective probability and subjective desirability or utility’

(Jeffrey, 1983, p. xi)

key assumption:

Agents’ actions maximise their expected utilities.

If we understand this as instrumental action, we run into a problem ...

habitual process

Action occurs in the presence of Stimulus.

Agent is rewarded [/punished]

Stimulus-Action Link is strengthened [/weakened] due to reward [/punishment]

Given Stimulus, will Action occur? It depends on the strength of the Stimulus-Action Link.

‘goal-directed’ process

Action leads to Outcome.
 

Belief in Action-Outcome link is strengthened.

Agent has a Desire for the Outcome
 

Will Action occur? It depends on the Belief in the Action-Outcome Link and Agent’s Desire.

Not all of the processes which influence action are processes that maximise expected utilties.
After all, there are habitual processes.

This book has ‘a philosophical end: elucidation of the notions of subjective probability and subjective desirability or utility’

(Jeffrey, 1983, p. xi)

key assumption:

Agents’ actions maximise their expected utilities.

objection:

The assumption is unjustified given the dual-process theory.

Note that formulating the objection properly is quite tricky.
One way to formulate it is to ask whether the notions of subjective probability and subjective desirability are supposed to be those which feature in the goal-directed process.
If they are, then the existence of habitual processes entails that the assumption that agents’ actions maximise their expected utilities is false.
If they are not, then we currently lack a way to identify psychological counterparts of the decision-theoretic notions, and so to transition from a merely descriptive model to a theory capable of explaining why people act.
So this is the objection. Agents’ actions do not actually maximise their expected utilties.
Is it a good objection?
Only use this if needed.

1. Instrumental actions maximise agents’ expected utilities.

2. Decision theory provides an ‘elucidation of the notions of subjective probability and subjective desirability or utility’ (Jeffrey, 1983, p. xi).

3. The notions elucidated are those of belief and desire, which also feature in the model of goal-directed processes.

4. Some instrumental actions are dominated by habitual processes.

5. Habitual and goal-directed processes can pull in opposing directions.

Key is that these claims undermine any basis for expecting the actual choices made to respect the axioms.

response 1

‘the laws of decision theory (or any other theory of rationality) are not empirical generalisations about all agents. What they do is define what is meant ... by being rational’

(Davidson, 1987, p. 43)

but: elucidation was our goal

This is maybe fine if you want to use decision theory like that.
But we wanted to use it to elucidate states which play a role in a theory that offers a causal explanation of action.
So this response really just conceeds the objection.
But this is important because it shows that ...

We are not objecting to decision theory.

We are objecting to a particular construal of it (as an elucidation).

construal 1: decision theory provides ‘mathematically complete principles which define “rational behavior”’

construal 2: decision theory provides an elucidation of belief and desire

NB: the claim is that decision theory provides an elucidation of belief and desire *as featured in goal-directed processes*. (So strictly speaking Jeffrey might have succeeded in elucidating the notions he was interested in.)
Interesting side question is whether the narrowly normative construal can stand. After all, it seems to require that what subjects are optimising is utilities specified in ways that do not reflect the beliefs and desires that feature in the goal-directed process.
I am not going to consider that further but I do think there is an opportunity for some original research on the basis for a normative construal of decision theory as defining rational behaviour.

response 2

It’s an approximation; the details don’t matter.

This is tempting because habitual processes will tend, when things to well, to maximise expected utilities.
We might even see game theory as characterising what the habitual processes are a shortcut to achieving.

but: prediction vs elucidation

This might be a fine response if we were using decision theory to predicting behaviour.
Sometimes you gain predictive power by abstracting from messy impelementation details.
And **seen from the outside**, agents will likely behave much as if they were maximising their expected utiltities.
But our aim was to elucidate the notions of belief and desire as they feature in an explanation of action. To do that, we have to identify them with the theoretical constructs of decision theory.
If agents’ actions do not conform to the axioms (transitivity of preferences etc), then they lack beliefs and desires. And so our theory fails to elucidate the things we wanted it to elucidate.
So this second response also really just conceeds the objection. It distracts us by pointing to other applications of decision theory. But we were not objecting that decision theory has no other applications. We were objecting to a specific application of it.

response 3

What maximises expected utility are not actions but goal-directed processes.

This is the most interesting line of response I know. It offers a substantial change to the construal of decision theory but also attempts to avoid the objection.

This book has ‘a philosophical end: elucidation of the notions of subjective probability and subjective desirability or utility’

(Jeffrey, 1983, p. xi)

key assumption:

Agents’ actions goal-directed processes maximise their expected utilities.

habitual process

Action occurs in the presence of Stimulus.

Agent is rewarded [/punished]

Stimulus-Action Link is strengthened [/weakened] due to reward [/punishment]

Given Stimulus, will Action occur? It depends on the strength of the Stimulus-Action Link.

‘goal-directed’ process

Action leads to Outcome.
 

Belief in Action-Outcome link is strengthened.

Agent has a Desire for the Outcome
 

Will Action occur? It depends on the Belief in the Action-Outcome Link and Agent’s Desire.

problem now is that preferences feature twice in the scheme of things!
Is the preference that dictates the reward the same or different?

Concerning the habitual process, what makes outcomes rewarding?

possibility 1:

the very system of preference that is involved in the goal-directed process

possibility 2:

not the system of preference that is involved in the goal-directed process

Each has potential problems.
possibility 2: now it looks like the two processes might pull in different directions. What if anything ensures that they are not fighting against each other in the way that people with different preferences might? (My habitual processes love bananas and hate chocolate ...)
possibility 1: if this is right, decision theory did not provide a full elucidation of the notion of preference. So we have discovered something substantial: decision theory provides at most a partical elucidation. And it is unclear how to develop the elucidation further.

response 4

distinguish computational theory from implementation details

[This is currently what I take to be the most promising response.]
Both habitual and goal-directed processes are supposed to maximise expected utility.
Each does so only within limits and so involves risks of error.
But the limits are a matter of the implementation (speed--accuracy trade-offs). They are not part of the computational description.
If this is right, there is no need to suppose that agents’ actions actually conform to the axioms. Just that they are in some sense supposed to do so, that this is an aim of the various processes involved in action selection.

How to implement a utility maximizing agent?

option 1: search through potential actions, imagine consequences, evaluate how good they’d be

option 2: estimate best action from past rewards

option 3: combine 1 and 2

Having two processes
allows you to make complementary
speed--accuracy trade-offs:
habitual processes are fast but limited, whereas goal-directed processes are more flexible but slower

Any broadly computational system will face ...

response 4

distinguish computational theory from implementation details

response 5

seek an alternative

‘Expected utility theory [...] has come under serious question [...]

There is now general agreement that the theory does not provide an adequate description of individual choice: a substantial body of evidence shows that decision makers systematically violate its basic tenets.

Many alternative models have been proposed’

(Tversky & Kahneman, 1992, p. 297)

response 6

they do not exist

‘The problem with measuring risk preferences is not that measurement is difficult and inaccurate; it is that there are no risk preferences to measure---there is simply no answer to how, ‘deep down’, we wish to balance risk and reward.

And, while we’re at it, the same goes for the way people trade off present against future; how altruistic we are and to whom; how far we display prejudice on gender, race, and so on ...

there is no point wondering which way of asking the question [...] will tell us what people really want.

there can be no method...that can conceivably answer this question, not because our mental motives, desires and preferences are impenetrable, but because they don’t exist

(Chater, 2018, pp. 123--4)

conclusion so far

This book has ‘a philosophical end: elucidation of the notions of subjective probability and subjective desirability or utility’

(Jeffrey, 1983, p. xi)

The dual-process theory of instrumental action, if true, complicates Jeffry’s claim that decision theory provides an elucidation of these notions.