Generating options and choosing between them rely on distinct forms of value representation

Abstract

Humans have a remarkable capacity for flexible planning, deliberating among actions by modeling their likely outcomes. This form of model-based planning allows us to adapt to the specific features of diverse circumstances. In real-world decision making, however, planning faces an important challenge: There are often an enormous number of possible actions to choose among, far too many for exhaustive consideration. There is a crucial, understudied “pre-planning” step in which, among myriad possibilities, a few good candidates come to mind with minimal effort. How do people accomplish this? We show that people use computationally frugal habits to propose a few candidate actions based on their general value across a range of contexts. Deliberative planning is then deployed just within this tractable set, updating value estimates based on context-specific features. This hybrid architecture combines the efficiency of habit and accuracy of planning, illuminating how valuable thoughts come naturally to mind during decision-making.

Habits and Goals in Human Behavior: Separate but Interacting Systems

“…From a goal systems perspective, as the number of possible means increases, the available activation for each specific means decreases, forcing people to select among potential courses of action (Kruglanski et al, 2002). In fact, the narrowing of response options may be an important function of habits, as they constrain the broad set of responses that people could give at any moment (Morris et al, 2019).…”

How We Know What Not To Think

“…We present a simple illustration of how value-guided consideration sets could be adaptive in decision making ( Figure I). This proposal is closely related to that in [113], which provides an algorithmic implementation of our computational proposal for a specific decision problem. When facing a decision with N options, people first generate a consideration set, C, of K << N options, where the probability of inclusion, i, of each option in C is proportional to the 'cached value' of that option [denoted CV(i)], which correlates imperfectly with the expected value of the option in the specific decision context.…”

“…Building on this approach, an optimality analysis was conducted [113], asking when it is optimal to sequence heuristic and deliberative processes (a 'hybrid' model) or instead to act either purely heuristically or purely deliberatively (two 'pure' models). Exploring a wide range of parameter settings, they find that the hybrid approach is often favored, pure habit is sometimes favored, and pure deliberation is rarely favored.…”

“…In sum, consideration sets could plausibly be constructed according to a readily available, computationally cheap representation of the value of an action 'in general', whereas actions are chosen from the consideration by computing online a representation of the value of an action 'specifically' for the present circumstance [83,89]. Initial evidence indeed suggests both a crucial role for cached values in the generation of the consideration set, and a separate role for task-specific values in the choice of a final option from the consideration set [113]. For instance, when choosing dinner for a guest with unusual tastes, people tend to generate potential dinners by sampling options that they themselves typically enjoy (relying on cached value representations), and then choose from within the options in this consideration set the particular dinner that best satisfy their guest's tastes (a task-specific value representation) [113].…”

“…Initial evidence indeed suggests both a crucial role for cached values in the generation of the consideration set, and a separate role for task-specific values in the choice of a final option from the consideration set [113]. For instance, when choosing dinner for a guest with unusual tastes, people tend to generate potential dinners by sampling options that they themselves typically enjoy (relying on cached value representations), and then choose from within the options in this consideration set the particular dinner that best satisfy their guest's tastes (a task-specific value representation) [113].…”