Mastering Machine Learning Algorithms

Environment

The environment is the entity where the agent has to reach its goals. For our purposes, a generic environment is a system that receives an input action, a_t (we use the index t because this is a natural time process), and outputs a tuple composed by a state, s_t+1, and a reward, r_t+1. These two elements are the only pieces of information provided to the agent to make its next decision. If we are working with an MDP and the sets of possible actions, A, and states, S, are discrete and finite, the problem is a defined finite MDP (in many continuous cases, it's possible to treat the problem as a finite MDP by discretizing the spaces). If there are final states, the task is called episodic and, in general, the goal is to reach a positive ...