CS & AI Student | Aspiring Software Developer | Passion for ML and Wolfram
Reinforcement learning (RL) is often considered the odd-one-out of the three subfields of AI, as its principles and methodology of using data is quite different from supervised and unsupervised learning.
The goal of RL is to develop agents (algorithms that can act) that make decisions based on interactions with an environment, is popularly defined as "a goal-driven approach to decision making problems".
Theoretically speaking however, RL can be best described as a data-driven extension of markov decision processes (MDPs), (see Tranditional AI).
In the RL paradigm, an agent can move between environmental states through actions and receive rewards based on the current state.
The goal of designing an RL algorithm is to create agents that maximise the sum of these rewards over time.
Despite the simple premise, this paradigm can be applied to a variety of problems, with different algorithms and techniques.
In many ways, RL is the most intuitive of the three subfields of ML, as we can generally make sense of the strategies produced by the algorithms, even as deep learning becomes more involved.