COGNIVITE SCIENCE · 2015-11-24 · Formation of expectancies, E, about the values of each option j on trial t. Specifically, the momentary utility u j (t) serves as input to a learning

COGNIVITESCIENCE

2015 인지과학 방법론

CogSci 발제

2015-20103

인지과학 협동과정

2학기

김지윤

Behavioral

Economics:

CO

NTEN

TS

CogSci:

Exploring Complexity in Decisions from Experience: Same Minds, Same Strategy

“지금 인류는 ‘호모 스키초프레닉(Homo Schizophrenic·정신분열증 환자)’입니다.

생각을 할 때 이쪽저쪽을 왔다 갔다 한다는 뜻입니다.

2008년 같은 위기에서는 전통 경제학에 나오는 ‘호모 이코노미쿠스

(Homo Economicus·인간은 합리적으로 판단하고 결정한다는 의미)’는 없습니다.

전통 경제학에는 중요한 게 빠져 있어요. 인간의 행동과 심리가 그것입니다.”

Behavio

ral

Econom

ics

Behavioral

Economics

분야에서

다뤄지는

주제들

Non-standard preferences

Self‐control problems

Reference dependent preferences

Risk preferences

Social preferences

Non‐standard beliefs

Overconfidence

Law of small numbers

Projection bias

Non‐standard decisions

Framing

Limited attention

Menu effects

Persuasion and social pressure

Emotions

Behavio

ral

Econom

ics

| Please take a piece of paper

| Pick any integer number between 0 and 100. And write it down on the paper.

| You will receive a prize, 5,000 KRW, if your number is closest to two thirds (2/3) of the average of all chosen numbers.

Behavio

ral

Econom

ics

Behavioral

Economics

Players who think others choose randomly guess the average is 50should choose 33 (= (2/3)*50).

If a player thinks others think that way, she should choose 22 = (2/3)*33.

A player with N = the number of iterations of reasoning shouldchoose

As N goes to infinity, the optimal number is 0.

“과연 인간은 합리적일까?

선택지가 달라지면 , 상황이 복잡해지면

그에 맞춰서 의사결정을 달리하게 될까?”

Exploring Complexity in Decisions from Experience:

Same Minds, Same Strategy

Emmanouil Konstantinidis, Nathaniel J. S. Ashby

& Cleotilde Gonzalez

Department of Social & Decision Sciences

Carnegie Mellon University, Pittsburgh, PA 15213, USA

CogSci

CogSci 2016

CogSci

Exploring Complexity in Decisions from Experience: Same Minds, Same Strategy

Research Question

Method

Participants

Task

Procedure

Behavioral Results

Modeling Structural Complexity

The model

Model evaluation

Modeling Results

Model fitting

Model parameters

Discussion

Not to “put all our eggs in one basket”

Ashby

/ Found support for the diversity hypothesis across all levels of complexity.

/ The diversity hypothesis: suggests that participants distribute their choices, choosing from options proportional to their experienced EVs.

/ This hypothesis was tested using an experimental design by increasing the number of options available to choose from.

| Research Question

| Method

| Behavioral Results

| Modeling Structural Complexity

| Modeling Results

| Discussion

Previous Research:C

ogSci

The purpose of the present work:

To decompose its underlying cognitive and psychological underpinnings by using computational modeling analysis.

Cognitive models:

To assess latent psychological factors that affect behavior in a task and to quantify these processes via model parameters.

The main hypothesis:

Diversification in choice originates from similar cognitive processes; in other words, model parameters will be similar across conditions and levels of structural complexity.

CogSci

Present Research:| Research Question

| Method



| Modeling Results

| Discussion

CogSci

Participants:

405 participants,

recruited from Amazon Mechanical Turk

(M age = 33.46, 47% female)

Task:

Participants had to make 200 consequential choices in one of four conditions: choices involving

| 2 (C2: N = 99)

| 4 (C4: N = 100)

| 8 (C8: N = 100)

| 16 (C16: N = 106)

In each condition,

there was an equal number of

| safe (two moderate outcomes with equal probability)

| risky (low probability of a high outcome, but higher probability of a low outcome)

| Research Question

| Method



| Modeling Results

| Discussion

CogSci

The maximizing option (i.e., the option with the highest EV) was option S1 across conditions.

After each choice, participants received feedback about the outcome of their decision.

Table 1: Safe (S) and risky (R) gamble pairs outcomes in points (OS1-OS2 and OR1-OR2), probabilities (50%-50% and 20%-80%), and expected values in points (EVS and EVR) in all experimental conditions (C2, C4, C8, C16).

| Research Question

| Method



| Modeling Results

| Discussion

CogSci

Procedure:

Participants presented with either 2, 4, 8, or 16 options (between subjects),

Play the options in order to learn what outcomes were possible as their outcomes and probabilities would not be provided.

They were told that their goal was to earn as many points as possible, as points would be converted to money (40 points = $0.01).

After participants made their 200 decisions, they were informed of their total earnings and thanked for their time.

| Research Question

| Method



| Modeling Results

| Discussion

CogSci

Examine the pattern of

choice selections across conditions:

The first step:

Examine the pattern of choice selections across conditions.

According to the diversity hypothesis,

Expect participants’ strategies to depart from maximization and to rather show a pattern where they allocate their choices based on the EV of each option

“몰빵 보다 분산”

“분산은 돌아오는 것이 큰 순서대로”

| Research Question

| Method



| Modeling Results

| Discussion

CogSci

Figure 1: Proportion of choices from each option across conditions.Numbers in parentheses indicate each option’s EV.

Direct mapping between the rank ordering of options based on their EV and the proportion of choices each of them receives.

| Research Question

| Method



| Modeling Results

| Discussion

CogSci

Passible Explanation:

“몰빵 보다 분산”

“분산은 돌아오는 것이 큰 순서대로”

왜?

The tendency to spread choices might be a result of extended exploration in the first blocks of the experiment and participants select more frequently the maximizing option in the last trials of the task.

“초기 탐색을 위해?”

확인해보자!

In order to examine this possibility, we tested whether the experienced outcome from each option (first 160 trials) is a significant determinant of choice selection in the last trials of the experiment (40 trials).

| Research Question

| Method



| Modeling Results

| Discussion

CogSci

Passible Explanation:

“초기 탐색을 위해?”

Conducted a mixed-effects multiple regression predicting the proportion of choices for each option in the last 40 trials by the average experienced outcome up to trial 160.

The analysis showed that each option’s average experienced outcome significantly predicts choice proportions in the last 40 trials, b = .01,z = 23.80,p < .001.

This result suggests that participants allocate their choices according to each option’s EV even in the later stages of the experiment and after extended feedback and interaction with the task.

“어라? 계속 분산???”

| Research Question

| Method



| Modeling Results

| Discussion

CogSci

“조건에 관계없이 사람들 계속 분산,

그 기저의 심리적, 인지적 과정은?”

We utilized a computational modeling analysis in order to examine whether the psychological and cognitive processes underlying choice behavior are similar across conditions.

“Computational modeling analysis 이용”

| Research Question

| Method



| Modeling Results

| Discussion

CogSci

The model:

“조건에 관계없이 사람들 계속 분산,

그 기저의 심리적, 인지적 과정은?”

Employed a reinforcement-learning (RL) model, which incorporates three basic assumptions regarding the decision process.

These assumptions reflect psychological and cognitive processes.

Assumption 1:

subjective evaluation of received feedback

Assumption 2:

Formation of expectancies

Assumption 3:

A choice

| Research Question

| Method



| Modeling Results

| Discussion

CogSci

The model:

Assumption 1:

Relates to the subjective evaluation of received feedback in each trial after selecting an option by a utility function.

Assumption 2: Assumption 3:

| Research Question

| Method



| Modeling Results

| Discussion

The free parameter α (0 ≤ α ≤ 1) determines the shape of the utility function. When α equals 1, the subjective utility matches the received payoff, whereas values less than 1 result in a curved utility function.

CogSci

The model:

Assumption 1: Assumption 2: Assumption 3:

| Research Question

| Method



| Modeling Results

| Discussion

Formation of expectancies, E, about the values of each option j on trial t. Specifically, the momentary utility u j (t) serves as input to a learning rule which updates these expectancies.

The free parameter A (0 ≤ A ≤ 1) indicates how much the old expectancy, E j (t −1), is modified by the prediction errorLarge values of A reflect rapid forgetting and strong recency effects, whereas smaller values of A indicate weak and recency effects and long associative memories.

CogSci

The model:

Assumption 1: Assumption 2: Assumption 3:

| Research Question

| Method



| Modeling Results

| Discussion

A choice, which is a probabilistic function of the relative strength of each option, is made.

The free parameter θ (sensitivity parameter) controls the degree to which choice probabilities match the formed expectancies.

where c ranges between 0 and 5, with values close to 0 indicating random choice and values close to 5 suggesting deterministic-exploitative choice.

CogSci

Model Fitting:

“우선, 모델이 참여자들의 선택을 잘 담고 있는지 검증”

Ensure that the model we employed provides a good fit to the data and an accurate representation of the participants’ choices.

| RL model

| Random pick model which assumes random choice on every trial

| Statistical mean-tracking model which assumes choice is based on the relative strength of each option’s mean observed payoff in each trial.

Schwartz Bayesian information criterion (BIC)

| Research Question

| Method



| Modeling Results

| Discussion

CogSci

Model Fitting:

“모델이 참여자들의 선택을 잘 담고 있는지 검증”

µBIC: the mean BIC

nBIC: the number of participants best fit by each model

Table 2: Model fitting results: Values of mean BIC (µBIC; lower values indicate better fit) and total number of participants best fit by each model (nBIC) across complexity conditions.

IRL model outperforms, across all conditions.

= Predictions regarding choice performance improve if we use the RL model.

| Research Question

| Method



| Modeling Results

| Discussion

CogSci

Model Fitting:

The observed mean proportion of choices from each option across 200 trials (Data) and model’s predictions (Model).

The model accurately predicts the rank order of each option

Figure 2: Mean proportion of observed choices (Data: Solid line) and mean predicted choice probabilities (Model: Dashed line) for each option across conditions (C2, C4, C8, and C16).

| Research Question

| Method



| Modeling Results

| Discussion

CogSci

Model Fitting:

The model’s predictions for the overall proportion of choices are also very close to the observed ones

“모델 검증의 의미＂

The accuracy of the model provides further support that the model we employed is a good representation of the factors responsible for observed behaviorFigure 3: Overall observed proportion of

choices from each option (Data) and overall predicted choice Probabilities (Model) across conditions. Numbers in parentheses indicate each option’s EV.

| Research Question

| Method



| Modeling Results

| Discussion

CogSci

Model Parameters:

According to the diversity hypothesis,

Do not maximize their overall winnings by consistently selecting the most profitable option,

but they prefer to distribute their choices by rank ordering the options based on their EV.

If this hypothesis stands true,

then model parameters should not differ across conditions, reflecting similar underlying psychological processes.

“어떤 상황에서도 분산하려는 인지적 과정

= 어떤 조건에서도 모델 동일”

| Research Question

| Method



| Modeling Results

| Discussion

CogSci

Model Parameters:

Kruskall-Wallis test, revealed | significant effects of condition in α, K(3)= 14.84, p=.002, and A parameters, K(3) = 18.96, p < .001,

| No differences in c parameter, K(3) = 4.95, p =.18.

Table 3: Mean (M) and median (Md) values of model parametersacross conditions.

| Research Question

| Method



| Modeling Results

| Discussion

Pairwise comparisons showed that differences across conditions found in parameters α and A are mainly because of condition C2 being significantly different from the remaining conditions. No differences were observed between the remaining conditions (C4, C8, and C16).

CogSci

Model Parameters:

Computational modelling analysis showed that condition C2 is different with participants

| showing higher updating rates (parameter A) compared to the other conditions. In other words, participants in C2 condition rely more on newly received payoffs and thus show, on average, stronger recency effects and rapid forgetting of the previously formed expectancies compared to participants in the remaining conditions.

| parameter α, closer look at the median value indicates that this may be due to some extreme cases.

| Research Question

| Method



| Modeling Results

| Discussion

| Research Question

| Method



| Modeling Results

| Discussion

CogSci

Model Parameters:

The way people behave in different levels of structural complexity.

Differences in parameter A between conditions is the result of a different pattern of behavior in the first 50 trials. Specifically, participants in condition C2 learn to pick the maximizing option faster than participants in other conditions

CogSci

Model Parameters:

Does not necessarily mean that participants in the C2 condition update more.

participants’ choice behavior stabilizes after the initial exploration phase and they do not update the expectancies of the options.

| Research Question

| Method



| Modeling Results

| Discussion “다른 option들 보다빨리 배운다는 것 뿐, 초기 탐색 후에 업데이트 일어나지 않음.”

CogSci

Summary:

The main objective:

구조적 복잡도가 다른 조건들에서의 선택 전략

Sought to answer whether an observed tendency to spread choices and favor diversity is characterized by similar cognitive mechanisms across different levels of structural complexity.

| Research Question

| Method



| Modeling Results

| Discussion

CogSci

Summary:

Two main implications:

| First, participants showed a tendency to distribute and diversify choices, despite the fact that they would move away from the optimal EV maximization option: choice proportions follow the rank ordering of options based on their EV

| Research Question

| Method



| Modeling Results

| Discussion

| Second, implication comes from computational modeling analysis. Inspection of model parameters across conditions suggests that the underlying psychological processes responsible for observed behavior are essentially identical across different levels of structural complexity.

Limitation :

Limitation 1:

The parameter estimates across conditions come from different participants| employing within-subjects designs

Limitation 2:

The RL model was compared against two rather “weak” and cognitive-free statistical models. | Other cognitive models

CogSci

| Research Question

| Method



| Modeling Results

| Discussion

Discussion :

Behavioral Economics Application

| Marketing

| Financial management

| Nudgehttp://news.chosun.com/site/data/html_dir/2009/09/11/2009091101231.html

CogSci

| Research Question

| Method



| Modeling Results

| Discussion

CogSci

Reference Ashby, N. J. S., Konstantinidis, E., & Gonzalez, C. (2015). Exploring Complexity in Decisions from Experience: Same Minds, Same Strategy. CogSci 2015

Ahn, W.-Y., Busemeyer, J. R., Wagenmakers, E.-J., & Stout, J. C. (2008). Comparison of decision learning models using the generalization criterion method. Cognitive Science, 32, 1376–1402.

Ashby, N. J. S., Konstantinidis, E., & Gonzalez, C. (2015). Too many baskets: A misguided preference for diversity in decisions-from-experience.

Kahneman, D., & Tversky, A. (1979). Prospect theory: Ananalysis of decision under risk. Econometrica, 47, 263–291.

Documents

COGNIVITE SCIENCE · 2015-11-24 · Formation of expectancies, E, about the values of each option j on trial t. Specifically, the momentary utility u j (t) serves as input to a learning