Performance-oriented selection of bounded LP objective directions

Investigate how to search among the feasible moment-matching directions for objective directions that maximize the performance of the resulting data-driven optimal-control policy.

Background

The boundedness conditions can admit many feasible directions when the combined number of observed and auxiliary samples exceeds the number of polynomial-feature coefficients. The paper normalizes the dual weights to remove scale ambiguity, but leaves unresolved the problem of selecting a particular feasible direction based on control performance rather than relying on an arbitrary solver-selected solution.