Kahneman, D., & Tversky, A. (1981). The simulation heuristic. Retrieved from http://www.dtic.mil/cgi-bin/GetTRDoc?AD=ADA099504

Summary

In this paper, Kahneman & Tversky discuss how people construct mental simulations of hypothetical scenarios, and in particular, counterfactual scenarios. They argue that people rely on a “simulation heuristic”, which is defined as constructing scenarios in order to estimate the probability of events.

In the introduction, Kahneman & Tversky provide a nice definition of what they mean by simulation (pg. 1-2):

Our starting point is a common introspection: there appear to be many situations in which questions about events are answered by an operation that resembles the running of a simulation model. The simulation can be constrained and controlled in several ways: the starting conditions for a ‘run’ can be left at their realistic default value, or modified to assume some special contingency; the outcomes can be left unsecified, or else a target state may be set, with the task of finding a path to that state from the initial conditions. A simulation does not necessarily produce a single story, which starts at the beginning and ends with a definite outcome. Rather, we construe the output of simulation as an assessment of the ease with which the model could produce different outcomes, given its initial conditions and operating parameters. Thus, we suggest that mental simulation yields a measure of the propensity of one’s model of the situation to generate various outcomes, much as the propensities of a statistical model can be assessed by Monte Carlo techniques.

They also give a few use cases that seem to involve simulation:

  1. Prediction
  2. Assessing the probability of a specified event
  3. Assessing conditioned probabilities
  4. Counterfactual assessments
  5. Assessments of causality

Methods

In this paper, they focus primarily on counterfactual assessments (i.e., “if only…” scenarios). The first type of task that they focus on is that involving regret. The example scenario is that of two people who miss their flight, but by different margins (5 minutes vs. 30 minutes), and ask participants to ask who feels worse. Almost universally people respond that the person who missed their flight by 5 minutes feels worse than the other person. Kahneman and Tversky argue that this is because it is easier to imagine making it to the airport 5 minutes earlier than it is to make it to the airport 30 minutes earlier.

The second type of task they focus on involves people actually producing alternate scenarios that could have prevented something happening (for example, someone dying in a car crash). There were two conditions; one in which the driver left work early (“time” version), and one in which they took an unusual route (“route” version). In the time version, participants came up with alterations to the scenario like “he should have left at a different time” (26%), “he should have crossed the intersection more quickly” (31%), or “the other driver shouldn’t have been driving” (29%). In the route version, the majority of participants said “he shouldn’t have taken the different route” (51%), with some also saying “he should have crossed the intersection more quickly” (22%) or “the other driver shouldn’t have been driving” (20%).

The route version involves an unusal element (the route), and correspondingly, participants are vastly more likely to undo this element in their counterfactual simulation than the participants in the time version. Kahneman and Tversky explain this as one of three types of changes that can occur:

  1. Downhill change (increases probability, decreases surprise)
  2. Uphill change (decreases probability, increases surprise)
  3. Horizontal change (arbitrary value is changed, no change in probability or surprise)

In particular, the change in the route version is an example of a downhill change; thus, Kahneman and Tversky argue that when constructing simulations, people are biased towards scenarios that are more probable and less surprising. Moreover, people are biased towards making alterations to the scenario that are related to the main object or character (e.g., the behavior of the original driver, and not the other driver).

Algorithm

n/a

Takeaways

Kahneman and Tversky characterize the simulation heuristic as being biased towards downhill changes. I wonder if this could be explained by something like a stochastic search in the space of scenarios (e.g. something like Monte-Carlo tree search), where people are trying to maximize something like the posterior probability of the scenario given the alternate outcome. Due to the prior, this would result in a bias towards less surprising scenarios (e.g. undoing the change in route). This bias isn’t necessarily a bad thing, though. The prior is a well-motivated component, because you need something to constrain the space of possible scenarios—the narrative structure of the situation still needs to be coherent, and the structure of the prior ensures that.