Completion vs Optimality: Policy Gradient in Long-Horizon Cumulative-Damage Problems

In three linesStudy of policy-gradient method failures in long-horizon decision problems with cumulative damage. Authors identify two orthogonal failure modes and propose decomposition separating completion (reaching terminal horizon) and optimality (matching dynamic programming). Experiments on bricklayer career (49 steps) and NBA forward career (20 seasons).

Summary generated by Claude — human-verified