Preferences

Preferences#

A preference is a target distribution: what the System expects, and so prefers, to observe or to be. Preferences are entries of SystemParams and are ordinary distributions.

  • obs_targets, by modality: a Categorical over a discrete modality’s levels, or a Gaussian or MixtureGaussian over a continuous reading.

  • state_targets, by state factor: a target over states. Under the name of a discrete control node, a Categorical is a fixed cost per level of that control.

  • terminal_targets, by state factor: a target scored once, at the end of the planning horizon.

A per-step target may be conditional on other variables, a control node among them (C_deps on the System), and a terminal target may be conditional on other state factors (H_deps). A mixture can be used to shape a target, for instance flat over a region and steep at its edge. How targets enter the quantity a planner minimises is described under Objectives.

Targets for steps of several ticks#

accumulated_target

A per-tick target as the conditional target of a step, one distribution per level.