| Popis: |
From foraging for food to choosing a career, many decisions in life involve multi-step planning: choices made early ondetermine which choices will become available later. How do people plan in such contexts? We present a spatial MazeSearch Task (MST), that resembles real-life spatial search tasks, where subjects search for a goal in partially observed environmentswith a reward placed randomly. MST requires choosing a search path through the environment that balancesthe probability of success against the costs of making each observation. We use this task to probe the underlying computationalmechanisms of human planning under uncertainty. Using computational modeling and human experiments weevaluate 4 computational models that plan ahead, and 4 myopic heuristics that choose the next observation one step at atime.We found that: (1) human decisions in MST are best explained by models that plan ahead, as opposed to myopic heuristics;(2) an optimal model of planning, which is based on optimizing Expected Utility alone, is the best-fitting model foronly a small number of subjects; most subjects were best explained by a planning model that modified Expected Utilityby a probability weighting function based on Prospect Theory, temporal discounting of future decision states, or both; (3)people showed substantial individual differences in planning strategies; out of the 8 evaluated strategies, 6 (all, exceptfor two heuristics) were fitted to at least some individuals.Our results show that probability weighting – the principle of overestimating of small and underestimating large probabilities,proposed by Prospect Theory to model one-shot monetary gambles – applies to human sequential decisionmakingin naturalistic spatial environments. Likewise, we show that temporal discounting – often used in ReinforcementLearning to achieve convergence of decision-state values computed under infinite planning horizons – can also be usedto model how humans limit their planning horizon within finite-horizon tasks. |
