Humans and other animals are able to flexibly select among internally generated goals and form plans to achieve them. Still, the neuronal and computational principles governing these abilities are incompletely known. In computational neuroscience, goal-directed decision-making has been linked to model-based methods of reinforcement learning, which use a model of the task to predict the outcome of possible courses of actions, and can select flexibly among them. In principle, this method permits planning optimal action sequences. However, model-based computations are prohibitive for large state spaces and several methods to simplify them have been proposed. In hierarchical reinforcement learning, temporal abstractions methods such as the Options framework permit splitting the search space by learning reusable macro-actions that achieve subgoals. In this article we offer a normative perspective on the role of subgoals and temporal abstractions in model-based computations. We hypothesize that the main role of subgoals is reducing the complexity of learning, inference, and control tasks by guiding the selection of more compact control programs. To explore this idea, we adopt a Bayesian formulation of model-based search: planning-as-inference. In the proposed method, subgoals and associated policies are selected via probabilistic inference using principles of descriptive complexity. We present preliminary results that show the suitability of the proposed method and discuss the links with brain circuits for goal and subgoal processing in prefrontal cortex
