Demystifying Deep Reinforcement Learning

Introduction: What is Deep Reinforcement Learning?

Deep reinforcement learning is an area of machine learning which models the process of decision-making in the environment. It is based on decision-making environments that are composed of an agent and a feedback loop, where the agent takes actions in order to maximize its reward. Deep reinforcement learning focuses on how to train these agents to make better decisions given their current state and environment, rather than just performing high-level tasks without any understanding of how it influences the future rewards.

Essentially, this method of Machine Learning allows agents to reap the rewards and punishments for their actions and to learn from past mistakes. It is different from other machine learning algorithms in that it works on a trial and error approach, rather than what we think of as the more traditional supervised learning where humans provide feedback.

How Does Deep Reinforcement Learning Work?

Demystifying Deep Reinforcement Learning: It is a subfield of machine learning that focuses on how to train artificial agents to act optimally. This is done by following an agent-environment loop in which the agent takes actions in the environment, its reward function is updated based on the action taken, and it takes another action with a probability determined by its new value.

File:Markov diagram v2.svg
Demystifying Deep Reinforcement Learning: How the reinforcement learning feedback loop works. Diagram by EBatlleP – Own work, CC BY-SA 4.0,

The goal of deep reinforcement learning is to explore the space of all possible policy functions and find one that achieves good performance on some task or set of tasks. To do this, it must be equipped with an algorithm that explores this space efficiently, which can be accomplished through methods such as Monte Carlo tree search or Deep deterministic policy gradients.

More on Deep Reinforcement Learning

Use cases and applications

The practical use cases for this type of machine learning are not as broad as the use cases for regular reinforcement learning yet. However, it can be practically used in many areas in the future: self-driving cars, advanced AIs playing chess games / any games, and speech recognition, just to name a few.

MIT introductory class on Deep Reinforcement Learning

A good resource to look into if you’re keen to dive deeper into the topic on an introductory level. Hopefully, this will help further demystify deep reinforcement learning for you.

MIT 6.S091 by Lex Friedman

Implication on the future of AI

AI is not just an industry of the future. It is already shaping our lives in ways we don’t always notice. The potential future use cases for AI are exciting and diverse, and they’ll affect a wide range of industries.

Deep reinforcement learning is a type of machine learning that enables AI to learn from trial and error, without being explicitly programmed on what to do or what to expect. Reinforcement Learning was initially used to train autonomous robots, but it can also be used in many other ways. In fact, it’s credited with the recent rise of AlphaGo – the world’s first self-learning Go-playing computer program.