|
@KJuechems | |||||
|
New preprint with Jan Balaguer, Bernhard Spitzer and @summerfieldlab now out on this topic: Why are our decisions sometimes “irrational”? psyarxiv.com/6yhwg/
Take the example below. Which lottery do you prefer in scenario 1? What about scenario 2? 1/12 pic.twitter.com/222cbY94xf
|
||||||
|
||||||
|
Keno Juechems
@KJuechems
|
27. sij |
|
The expected values were simply rescaled by a constant factor between scenarios, so a rational agent should prefer the same option in each scenario. However, most people prefer A in scenario 1, but B in scenario 2. This is an example of an irrational preference reversal. 2/12
|
||
|
|
||
|
Keno Juechems
@KJuechems
|
27. sij |
|
Economic models can account for this phenomenon if subjective estimates of value and probability are nonlinear transforms of their objective counterparts. 3/12
|
||
|
|
||
|
Keno Juechems
@KJuechems
|
27. sij |
|
For example, a Nobel-prize winning economic theory proposes that the probability and value functions take the form shown below. 4/12 pic.twitter.com/FJxrJfBG8x
|
||
|
|
||
|
Keno Juechems
@KJuechems
|
27. sij |
|
Descriptively, these functions fit empirical data well. But why should humans have evolved such peculiar, non-linear functions for representing probability and value? Here, we address this question as an optimization problem. 5/12
|
||
|
|
||
|
Keno Juechems
@KJuechems
|
27. sij |
|
We ask under what constraints these canonical functions would be optimal. We start with two assumptions: (i) that humans wish to maximise their expected value, and (ii) that they have only finite computational precision, i.e. that decisions are noisy. 6/12
|
||
|
|
||
|
Keno Juechems
@KJuechems
|
27. sij |
|
Making these assumptions alone reproduces the canonical subjective value function. In other words, this nonlinear function is optimal for noisy agents. However, one additional assumption is necessary to also reproduce the probability function…7/12
|
||
|
|
||
|
Keno Juechems
@KJuechems
|
27. sij |
|
We assume that humans seek to maximize reward whilst minimizing uncertainty and incorporate this into our optimization via the entropy of potential outcomes of a lottery. Thus equipped, our simulations fit empirical data quite well: 8/12 pic.twitter.com/zg7hUtifJM
|
||
|
|
||
|
Keno Juechems
@KJuechems
|
27. sij |
|
Our simulations make two key predictions: i) that human choices should fall within the optimal range for capacity limited agents and ii) that the probability weighting depends on feedback certainty. 9/12
|
||
|
|
||
|
Keno Juechems
@KJuechems
|
27. sij |
|
In two experiments, we asked human participants to choose between two lotteries. We manipulated whether, after each choice, they received the expected value of their chosen lottery (certain feedback) or whether the lottery played out (uncertain feedback). 10/12
|
||
|
|
||
|
Keno Juechems
@KJuechems
|
27. sij |
|
Our results confirmed that participants adapted optimally to this manipulation and that only the uncertain group exhibited the canonical probability weighting function: 11/12 pic.twitter.com/S0ieaCyAYJ
|
||
|
|
||
|
Keno Juechems
@KJuechems
|
27. sij |
|
Taken together, our results offer a new perspective on human subjective utility and probability functions: that they represent an optimal adaptation to computational constraints imposed by biology. 12/12
|
||
|
|
||