Abstract
We discuss the problem of policy representation in stochastic
and partially observable systems, and address the case where
the policy is a hidden parameter of the planning problem. We
propose an adaptation of the Predictive State Representations
(PSRs) to this problem by introducing tests (sequences of actions
and observations) on policies. The new model, called
the Predictive Policy Representations (PPRs), is potentially
more compact than the other representations, such as decision
trees or Finite-State Controllers (FSCs). In this paper,
we show how PPRs can be used to improve the performances
of a point-based algorithm for DEC-POMDP.
