One of the problems with Dynamic Programming algorithms is the need for a full knowledge of the environment in terms of states and transition probabilities. Unfortunately, there are many cases where these pieces of information are unknown before the direct experience. In particular, the states can be discovered by letting the agent explore the environment, but the transition probabilities require us to count the number of transitions to a certain state and this is often impossible.
Moreover, an environment with absorbing states can prevent visiting many states if the agent has learned a good initial policy. For example, in a game, which can be described as an episodic MDP, the agent discovers the environment while learning ...
