Model-free and model-based training

RL algorithms that do not learn a model of how the environment works are called model-free algorithms. By contrast, if a model of the environment is constructed, then the algorithm is called model-based. In general, if value (V) or action-value (Q) functions are used to evaluate the performance, they are called model-free algorithms as no specific model of the environment is used. On the other hand, if you build a model of how the environment transitions from one state to another or determines how many rewards the agent will receive from the environment via a model, then they are called model-based algorithms. 

In model-free algorithms, as aforementioned, we do not construct a model of the environment. Thus, the agent has to take an action at a state to figure out if it is a good or a bad choice. In model-based RL, an approximate model of the environment is learned; either jointly learned along with the policy, or learned a priori. This model of the environment is used to make decisions, as well as to train the policy. We will learn more about both classes of RL algorithms in later chapters.