Negative Dynamic Programming II: With Non-Positive Rewards on Borel Spaces, the Optimal Return Satisfies the Optimality EquationResearch Paper
Negative dynamic programming is the study of infinite-horizon sequential decision problems in which every one-stage reward is non-positive and nothing is discounted. Equivalently, a non-negative cost accumulates forever and the controller minimizes its expected total. Stochastic shortest-path problems, optimal stopping with a cost per step, and inventory or…