Machine Learning

Reinforcement Learning

Reinforcement learning (RL) is a subfield of Machine Learning associated with how an intelligent Agent takes some Actions in an Environment, thereby transitioning through States, while seeking a Reward.

The goal of reinforcement learning is for the Agent to learn a Policy (decision making process, or collection of probabilities of taking an action ‘a’ while in state ‘s’.) that maximizes its Rewards over time.

Policy Gradient Methods

Part of the solution of the RL problem is to search the space of possible Policies when choosing the most suitable one. This can be large or infinite - so usually some proxy or surrogate of the space is needed, i.e. a stochastic optimization problem. Policy gradient methods employ gradient ascent techniques in the search.