Rollout Models of Human Planning with Exact Action Probabilities

Planning often involves mentally simulating (“rolling out”) possible futures before acting. Rollout‑based models of planning have seen great success in artificial intelligence, yet cognitive process models of human rollout-based planning remain scarce because their likelihoods are typically intractable. The only way to fit models with intractable likelihoods is with approximations which prevent accurate model fits, resulting in noisy estimates of latent planning parameters. We introduce Rollouts with Stochastic Early Termination and Tractable Estimation (ROSETTE), a class of process models that yields closed‑form action probabilities, enabling exact, simulation‑free parameter estimation. ROSETTE treats planning as an absorbing Markov chain where in a simulated state, the agent decides to either continue or terminate the rollout. Once the rollout is terminated, the agent decides whether enough information has been gathered to take an action, or whether a new rollout should be initiated. Because ROSETTE provides exact likelihoods, parameter estimates and model evidence are uncontaminated by sampling noise. We develop a first instantiation called Progress‑Regress ROSETTE (PROSETTE), in which rollouts terminate when the agent perceives sufficient progress toward or regress away from the goal. We fit PROSETTE to data from human participants in a multi-step planning task and found that it well accounted for a rich set of summary statistics of behavior. Our work sets the stage for testing rollout-based planning models as a promising category of models of human planning across a wide range of tasks.