Fault-Tolerant Control of Degrading Systems with On-Policy Reinforcement Learning