Материал: искусственный интеллект

Author Index

Table 1

Tversky and Shafir (1992) observed average two-stage gamble probabilities for Known outcome and Unknown outcome conditions. Within-participants version: Win condition, followed by the Lose condition and finally the Unknown condition with respectively 7 and 10 days in between. The between-participants version shows closely matching results.

Busemeyer, 2011; Busemeyer & Bruza, 2012). Our focus will be on a choice between risky prospects in a controlled environment, namely the two-staged gamble paradigm of the seminal paper of Tversky and Shafir (1992). In that paradigm each participant is exposed to three conditions of a choice between risky prospects. In each of the three scenarios the gamble was the same: even odds to win $200 or lose $100. Tversky and Shafir offered the participant to take the gamble or stop it, just after having taken that identical gamble. More specifically, they proposed three scenarios for that repeated gamble: one with a Win outcome of the initial gamble, one with a Lose outcome of the initial gamble, and a third scenario in which the outcome of the initial gamble was not disclosed. In this third scenario the initial stage gamble outcome was therefore Unknown to the participant. Their study fascinatingly showed that the probability of taking the second gamble was high at .69 and .59 when the first-stage gamble had been Won, respectively Lost, while a significant depression to.36 was observed for the probability when the outcome of the first-stage gamble was Unknown (Table 1). Tversky and Shafir argued their participants were inclined to take the second-stage gamble under Win condition, and also under Lose condition, but would not take the second-stage gamble under Unknown previous outcome condition. They reasoned this was a violation of the Sure Thing Principle, and named the occurrence of this depressed gamble probability in the Unknown previous outcome condition a Disjunction Effect. Since their original accomplishment, the corroboration of the Sure Thing Principle in the twostage gamble paradigm has been rather uncertain. Kühberger, Komunska, and Perner (2001) replicated the two-stage gamble experiment but found no significant indication of the Disjunction Effect, neither when using the original payoff scheme (the equivalent in Austrian Shilling), nor when using lower payoffs (both executed in between-participants design), and nor for lowered payoff in a within-participants design, nor when real monetary payoff was used. Later Lambdin and Burdsal (2007) reported that the outcome results on the Disjunction Effect by Kühberger et al. (2001) may still be significant, by arguing that only the within-participants design should be considered relevant and that only those participants that play on Win and play on Lose should be retained to test the Sure Thing Principle. We are critical of the latter claim since it erroneously identifies such outcome patterns as deterministic cognitive responses to the proposed gamble instead of stochastically realised outcome responses (Section 1.1). To clarify our discussion of the relation between the Sure Thing Principle (STP), the Disjunction Effect (DE) and the Law of Total Probability (LTP) we expound the definition for each explicitly.

First, the Sure Thing Principle (STP) is the rational rule that a decision that is made under the assumption of either of two mutually exclusive and exhaustive events should then also be made when no information is available on these complementary situations. In particular, winning or losing the gamble are the disjoint and exhaustive outcomes that make up the full event space of the first-stage gamble.

Second, the Law of Total Probability (LTP) expresses the principle that the marginal probability of an event is the weighted average of the conditional probabilities of the event on disjoint conditions. Again, since winning or losing the initial-stage gamble make up all the disjoint conditions, the marginal gamble probability satisfies

where the weights for both disjoint conditions satisfy p (W ) + p (L) = 1. One needs to heed that the cognitive representation of the Unknown condition may not fully conform to the formal disjunction W or L, of the cognitive Win representation and the Lose representation. The semantics of the chosen formulation of the Unknown condition could cause the observed conditional probability relations to be at variance with the Law of Total Probability, Eq. (1). Due to this issue, in our study we have meticulously formulated the description of the Unknown condition in the two-stage gamble, Section 2, to express it exactly as the disjunction, ‘or’, of the Win and Lose condition. In this article every mention of the ‘violation of the LTP’ should more carefully be understood as the ‘violation of a prediction derived from the LTP’.

Third, the Disjunction Effect (DE) is a statistical pattern in the empirical conditional gamble probabilities. The Disjunction Effect

The Disjunction effect is essentially the finding that the marginal probability p (gamble), which is formally equal to p (gamble W or L), is not bounded between the conditional probabilities, p (gamble W ) and p (gamble L). Therefore the Disjunction Effect is a violation of a prediction derived from the Law of Total Probability.

Notice that Tversky and Shafir considered the Disjunction Effect to be the statistical pattern in which a majority of participants

2

Effect, Eq. (2), more generally as any deflative violation of the LTP. Our definition of the DE encompasses the specification by Tversky and Shafir, while allowing us to use the designation for the exact same statistical anomaly in empirical gamble probabilities occurring for any particular payoff size – as we will show to be the case in our observations.

It is clear that the Sure Thing Principle and the Disjunction Effect are phenomena that are not necessarily related, as was pointed out by Lambdin and Burdsal (2007) and which we will further discuss in Section 1.1.

One key objective of the present work is to provide a thorough examination of the Disjunction Effect that resolves the above ambiguities, partly by extending the original paradigm by incorporating factors which might offer insight regarding the inconsistencies in related empirical work. Notably, we aim to examine whether the Disjunction Effect might be dependent, (i) on the size of the gamble payoff, (ii) on the order of presentation of the second-stage gambles with Win, Lose and Unknown previous outcome conditions, (iii) on the context brought about by how second-stage gambles are conditioned in a particular block and, (iv) on the risk attitude of the participant. In order to realize these additional manipulations, as well as ensure variability regarding the pertinent individual characteristics of participants, we decided for a large sample, online implementation of our main experiment (which follows a smaller exploratory one). This type of approach – using ‘Mturk’ (Amazon Mechanical Turk) and a short time constraint on the experiment – required us to use a design with a small number of choices provided by each participant, so that an emphasis on analysis by grouped data was built-in into our method.1 However a partitioning of participants by risk attitude (and by set of played gamble patterns, SM 4) will still allow a more granular analysis of individual differences among participants.

In brief, our study did find a Disjunction Effect which depends on the risk attitude of the participant and on the order of presentation of the two-stage gambles with Known or Unknown previous outcome conditions, Section 3. Remarkably, for more risk averse participants we even found that the direction in which the prediction derived from the law of total probability is violated is directly related to the order of Known outcome and Unknown outcome second-stage gambles. No evidence however was found for the violation of the Sure Thing Principle itself.

From a theoretical perspective we found that the model of Tversky and Shafir (1992) based on Prospect Theory – the original explanation they offered for the Disjunciton Effect – cannot properly explain the observed data from our present study, Section 4 (and Appendix B). Previously empirical observations with ‘non-classical’ probability structure – as in the present two-stage gamble experiment – have been modeled using quantum probability theory (Pothos & Busemeyer, 2009; Accardi, Khrennikov, & Masanori, 2009). But neither of these models for the two-stage gamble paradigm covered the order effects which we observed at present.

We address this theoretical challenge by developing two stochastic models, premised on an assumption that any choices are probabilistic functions of latent, dynamically evolving belief-action states informed by outcome conditions and payoff values. Both these models are decision process models, one based on the classical framework of Markov dynamics and the other on quantum dynamics.

The Markov dynamical process model we employ is based on the Kolmogorov differential equation to describe the dynamics and adheres to classical probability theory (Sonnenberg & Beck, 1993; Busemeyer, Wang, & Lambert-Mogiliansky, 2009). This is formally related to a diffusion model (for example, Ratcliff’s model, Smith & Ratcliff (2015)). A diffusion model is also a Markov process which uses the Kolmogorov differential equation. The difference between our Markov model and a diffusion model is that we are using a discrete state Markov model and the diffusion model is a continuous state model. Both models finally produce a predicted probability of making a response. We use that predicted probability to calculate the likelihood of the observed response (Section 4.2). Note, empirical results – superficially at least – at odds with the prediction derived from the Law of Total Probability may undermine our expectations of the suitability of a classical model. Nevertheless it is possible that violations of the predictions derived from the Law of Total Probability could be classically accounted for in a Markov model – for example through the inclusion of noise in the mapping between belief-action states and decisions.

The quantum model was designed in an analogous way to the Markov model. It is also a dynamic process model. Its dynamics is based on the Schrödinger differential equation. A quantum-like model is motivated by the observation of classically ‘irrational’ choice behaviour. A quantum-like model is based on logical rules alternative to those from Boolean logic, allowing for events to be order dependent (noncommutative) and context dependent (non-distributive). In full deployment, quantum-like models are both probabilistic and dynamic, and as such can reflect the stochastic process of decision making, Section 4.3.

For both the Markov and the quantum model, dynamical processes were driven by mathematical objects embodying parameters relevant to the psychology of the problem (e.g., heuristic utility, contextuality, carry-over and belief mixing), which produce a state of probabilities for different choices. Crucially, the quantum model represents an alternative philosophy in how the dynamics of a decision evolve: with a Markov model, at any given point in time there is a definite value to the internal state that will ultimately drive the response. With the quantum model it is impossible to assign a specific value to the internal state, prior to a response. That is, the response ‘collapses’ the inherent uncertainty in the quantum state; in quantum theory, a decision/response/measurement brings potentiality into being. This special kind of quantum uncertainty is the key characteristic of the theory which leads to the emergence

1 A Bayesian cognitive model based on the idea that each participant has a prior probability of gambling (which can change from experienced wins or losses) is complicated by the fact that a simple updating scheme cannot cover the observed inflation and deflation of probabilities in the Unknown condition nor the order effect. Should an updating scheme provide a change of a prior probability to gamble depending on the first-stage outcome condition - up on W, down on L and neutral on U - then a net approximately zero effect would result over the block of Known-outcome gambles. No deflation nor inflation of the Unknown-outcome gamble probability would therefore result from this updating scheme. Moreover, a basic Bayesian cognitive model does not a priori offer an implementation for order effects, and requires supplementary events describing the order of choices (Trueblood & Busemeyer, 2011).

3

Table 2

Observed gamble patterns with notation order Win-Lose-Unknown in two-stage gamble experiments. TS92 is the original Tversky and Shafir study with consecutive delayed presentation of conditions(Tversky & Shafir, 1992). KKP-Exp3 is the within-participants (‘third’) experiment of Kühberger, Komunska, and Perner with random order juxtaposed conditions (Kühberger et al., 2001). LB-Coin is the coin experiment of Lambdin and Burdsal with random order juxtaposed conditions (Lambdin & Burdsal, 2007). BBP-Exp1 is extracted from our Experiment 1 study for payoff parameter

X = 2, with randomly ordered conditions. BBP Exp2KtoU are the extracted X = 2 results from our Experiment 2 study with blocked Win and Lose outcome conditioned gambles preceding the block of Unknown outcome gambles. BBP Exp2UtoK has the blocked conditions in reverse order.

Dominant patterns are tagged with an asterisk.

of interference effects (see also, Busemeyer, Wang, & Townsend (2006)).

For both the Markov and the quantum-like models, we solve the dynamical differential equations up to a certain time point to compute the predicted probability of a response at that time point, which can be used to calculate the likelihood of the observed responses. Note, we did not collect response times. In past work on the disjunction effect, no one has looked at response times. Modeling data on response times is something to look at in future research. As a final note, regarding how a response is generated from predicted probabilities, we assume that if the model produces a state which indicates e.g. p (gamble) = .7, a response is sampled by a Bernoulli process producing either of the responses ‘gamble’ or ‘stop’ with probabilities .7 and .3 respectively. Therefore, no additional mechanism is required to go from probabilities to responses and model probabilities translate directly to predictions about choice proportions. Note, diffusion models are always applied to choice proportions in exactly the same way.

Finally, a third model is included as a comparison baseline to the dynamical process models, Section 4.4. The baseline model is essentially a Logistic regression which, as in latent trait modeling, will relate heuristic utility to theoretical gamble probabilities and which is further parametrized to mimic contextual and carry-over features shared by the two dynamical decisions models, Section 4.4.

1.1. The Sure Thing principle, the Disjunction Effect and the Law of Total Probability

The Disjunction Effect and the violation of the Sure Thing Principle are not necessarily related. In principle the aggregation of individual gamble strategies could mask the violation of the Sure Thing principle, and vice versa, a dominant strategy of violation of the Sure Thing principle should not necessarily lead to a Disjunction Effect in the aggregate data. Tversky and Shafir asserted a strong relation between the occurrence of the violation of the Sure Thing principle and the Disjunction effect. In support of this they observed that when the Disjunction Effect occurred the dominant gamble pattern followed by their participants was ‘gamble - gamble - stop’, where the notation order of the decisions always follows the fixed order ‘Win - Lose - Unknown’ of the conditions. We will use a letter notation to denote these choices as (g|W , g|L, s|U), where ‘g’ stands for ‘do the gamble’ while ‘s’ stands for ‘stop the gamble’. In this notation adhering to the STP by replying ‘gamble - gamble - gamble’ corresponds to the pattern (g|W , g|L, g|U ). In their study precisely 26 out of 98 participants (Table 2, col. TS92) violated the STP by performing pattern (g|W , g|L, s|U), while by the same token 14 out of 98 abided to the STP by performing (g|W , g|L, g|U ). Assessing the validity of the STP by a frequency comparison treats each outcome gamble pattern as composed of three logically correlated – consequentially evaluated– rational decisions. We assert that each gamble decision, hence each gamble triplet, should be considered a stochastic realisation of a latent trait.

What choice patterns then lie at the origin of the Disjunction Effect? In terms of gamble patterns, how does a violation of the Law of Total Probability come about? Strictly the Law of Total probability can be violated in two ways, the gamble probability under Unknown outcome condition either undershoots or overshoots the span of the probabilities for gambling under Known outcome conditions.

The occurrence of these violations is related to the frequency of specific gamble patterns.2 With three conditions in the two-stage gamble paradigm there are eight possible gamble patterns in total, Fig. 1. Each conditional gamble probability can be expressed as the marginal of the joint probabilities identified by their corresponding gamble patterns

2 A single binary response pattern to gambles neither indicates a violation of nor consistency with the LTP. The intrinsic probability – the individual’s probability to gamble, given some conditions (e.g., payoffs) – generates binary random variable responses. That is, a priori any individual pattern is not a reliable indicator with respect to putative violations of the LTP nor the STP. The intrinsic probabilities produce binary choices that lead to choice proportions when averaged across participants and a law of large numbers assures the convergence of observed choice proportions to the intrinsic probabilities.

4