Exploiting Similarity Information in Reinforcement Learning - Similarity Models for Multi-Armed Bandits and MDPs