Happy new year and welcome to a new decade of TDS. With the start of the new decade, it is worth looking back at the not so recent past. While most of the theory of data science and machine learning has been around since the 1960s, it was the availability of "big data" starting in the early 2000s which fueled the unparalleled growth and success of data science and machine learning across every major industry.
Today, this milestone is already 20 years in the past and, until the mid-2010s, the majority of progress was in "classical" machine learning which was able to achieve good results with a few 100 to 10,000 training examples. However, starting in the mid 2010s, an explosion in the size of datasets paved the way for deep learning models which were able to significantly outperform the previous generation. But since yesterday, even deep learning has this 2010ish vibe. So what comes next?
Clearly, it must be something that requires even more data. Maybe infinite data? Say hello to Reinforcement Learning – the new kid on the block. Reinforcement learning is probably the hottest candidate for achieving Artificial General Intelligence (AGI) but its hunger for data has held it back so far. It is no coincide that reinforcement learning has mainly been applied to scenarios that can generate nearly infinite amounts of data such as games or robotics.
Sounds cool – but what is reinforcement learning anyway and why should I care? Well, you came to the right place, we assembled a list of eight great articles to get you started on your journey into reinforcement learning. The first articles cover the basics, then we present some tutorials and hands-on articles and finally, there are some advanced topics for anyone that is already (or just became) a reinforcement learning pro.
Anton Muehlemann— Editorial Associate / AMLE at Determined.ai
Reinforcement Learning 101
By Shweta Bhatt – 6 min read
The perfect place to start your reinforcement learning (RL) journey. This article covers 5 basic RL questions.
DeepMind Unveils MuZero, a New Agent that Mastered Chess, Shogi, Atari and Go Without Knowing the Rules
By Jesus Rodriguez— 6 min read
This article reviews MuZero, an AI agent that mastered several strategy games by learning the rules from scratch. MuZero is the successor of the famous AlphaZero engine which beat AlphaGo which itself was the first AI to beat a professional Go player – a task that had long been seen as unachievable by any AI.
Reinforcement Learning for Real Life Planning Problems
By Sterling Osborne, PhD Researcher— 15 min read
A comprehensive guide on using RL to solve a real-world problem in which data is limited.
Reinforcement Learning Concept on Cart-Pole with Deep Q-Networks
By Vitou Phy— 6 min read
One of the prime examples of RL is the cart-pole problem. In this game, you try to balance a vertical pole upright by moving it left and right. This article shows how to solve this problem using Deep Q-Networks.
Advanced DQNs: Playing Pac-man with Deep Reinforcement Learning
By Jake Grigsby— 22 min read
This article presents improvements over classical DQNs and shows how it can be applied to become better and better at playing Pac-man.
Trade and Invest Smarter – The Reinforcement Learning Way
By Adam King— 19 min read
Tired of games? What about making some money? While we can’t guarantee any financial gains – you will definitely gain knowledge learning about TensorTrade – an open-source python framework for trading and investing using RL.
Everything you need to know about Google’s new PlaNet reinforcement learning network
By Cecelia Shao— 8 min read
This article reviews Google AI’s Deep Planning Network for RL. PlaNet addresses several shortcomings of classical RL models through a latent dynamics model, model-based planning and transfer learning.
Neural Architecture Search – Limitations and Extensions
By Alex Adam— 11 min read
Neural Architecture Search is the idea that instead of having an expert find a good model for a given problem, an RL algorithm does it for you. A major problem is the vastness of the resulting search and thus NAS algorithms take an absurd amount of computational resources. This article highlights these limitations and shows some possible extensions to make NAS perform better.
We also thank all the great new writers who joined us recently, Andrew Hao, Louis de Benoist, Hause, Dr. Robert Kübler, Nick Hallmark, Roberto Sannazzaro, Andy Chen, Dan Segal, Phoebe Wong, Martin Beck, Lydia Guarino, Sijo Vm, Zafiris Bampos, Paul-Ambroise Duquenne, Mohamed-Achref Maiza, Emma Sheridan, and many others. We invite you to take a look at their profiles and check out their work.