Meta-decision-making in hierarchical planning

Long-term planning in complex uncertain environments is computationally demanding and can significantly exceed an agent's information processing capacities. Hierarchical planning offers an elegant solution by dividing problems up and conquering them through distributing computation across state spaces and time. However, hierarchical representational abstraction comes at a cost—it can discard information necessary for optimal decisions, reflecting a fundamental trade-off between computational efficiency and solution quality. We formalized the adaptive meta-control of this trade-off using a meta-decision-making framework. Applied to a novel navigation task, our model predicts the circumstances under which the benefits of hierarchical planning outweigh the costs. We confirmed our predictions by means of two preregistered human experiments: participants adaptively shifted toward hierarchical planning when (1) the task space was larger, (2) the task horizon was longer, (3) available computational resources were limited, and (4) the environment contained uncertainty. This adaptive adjustment of planning strategy was manifest not only in initial planning choices, but also in the temporal distribution of computations, evidenced by reduced decision quality and longer reaction times at subtask boundaries. The adjustment was accurately captured by our meta-decision-making model. Our findings indicate how humans dynamically adapt their planning strategies in response to environmental demands, uncovering the resource-rational principles governing adaptive hierarchical abstraction in human planning.