Vision meets Reinforcement Learning
by observing training samples sequentially, rather than in batch
by hill-climbing, while estimating the underlying distribution --- PALO... and other ways to learn effective performance systems
by exploiting domain and context information
(Syncrude)