Search

It has been suggested that humans mentally simulate the outcomes of their actions when making decisions. However, this process can be challenging in real-world decision-making, which typically involves temporally extended decision trees with numerous potential outcomes. Here, we demonstrate with a computational model that temporally extended decisions can be achieved with just a few forward simulations, formalized as rollouts. We also show that, under resource constraints, performing many partial (shallow) rollouts can yield more favorable outcomes than performing fewer full (deep) rollouts. Additionally, our model captures behaviors traditionally attributed to pruning or satisficing strategies without the need for explicit heuristics, providing an alternative explanation for these phenomena. Finally, we show that the dynamics of value estimation over successive rollouts closely resemble evidence accumulation models. Our framework offers a plausible mechanism for temporally extended decision-making and provides insights into the neural underpinnings underlying this process.