Scaling Learning based Policy Optimization for Temporal Tasks via Dropout