Introducing Deep Reinforcement Learning

Yuxi Li
Yuxi Li
Oct 16, 2018 · 4 min read

The manuscript of Deep Reinforcement Learning is available now! It makes significant improvements to Deep Reinforcement Learning: An Overview, which has received 100+ citations, by extending its latest version more than one year ago from 70 pages to 150 pages.

It draws a big picture of deep reinforcement learning (RL) with many details. It covers contemporary work in historical contexts. It endeavours to answer the following questions: 1) Why deep? 2) What is the state of the art? and, 3) What are the issues, and potential solutions? It attempts to help those who want to get more familiar with deep RL, and to serve as a reference for people interested in this fascinating area, like professors, researchers, students, engineers, managers, investors, etc. Shortcomings and mistakes are inevitable; comments and criticisms are welcome.

The manuscript introduces AI, machine learning, and deep learning briefly, and provides a mini tutorial for reinforcement learning. The following figure illustrates relationships among these concepts, with major contents for machine learning and AI .Deep reinforcement learning is reinforcement learning integrated with deep learning, or deep artificial neural networks. A blog is dedicated to Resources for Deep Reinforcement Learning.

The manuscript covers six core elements: value function, policy, reward, model, exploration vs. exploitation, and representation; six important mechanisms: attention and memory, unsupervised learning, hierarchical RL, multi-agent RL, relational RL, and learning to learn; and twelve applications: games, robotics, natural language processing (NLP), computer vision, finance, business management, healthcare, education, energy, transportation, computer systems, and, science, engineering, and art.

Deep reinforcement learning has made exceptional achievements, e.g., DQN applying to Atari games ignited this wave of deep RL, and AlphaGo (Zero) and DeepStack set landmarks for AI. Deep RL has many newly invented algorithms/architectures, e.g., DQN, A3C, TRPO, PPO, DDPG, Trust-PCL, GPS, UNREAL, DNC, etc. Moreover, deep RL has been enjoying very abound and diverse applications, e.g., Capture the Flag, Dota 2, StarCraft II, robotics, character animation, conversational AI, neural architecture design (AutoML), data center cooling, recommender systems, data augmentation, model compression, combinatorial optimization, program synthesis, theorem proving, medical imaging, music, and chemical retrosynthesis, so on and so forth. A blog is dedicated to Reinforcement Learning applications.

In general, RL is probably helpful, if a problem can be regarded as or transformed to a sequential decision making problem, and states, actions, maybe rewards, can be constructed; sometimes the problem may not appear as an RL problem on the surface. Roughly speaking, if a task involves some manual designed “strategy”, then there is a chance for reinforcement learning to help. Creativity would push the frontiers of deep RL further with respect to core elements, important mechanisms, and applications.

Albeit being so successful, deep RL encounters many issues, like credit assignment, sparse reward, sample efficiency, instability, divergence, interpretability, safety, etc.; even reproducibility is an issue.

Six research directions are proposed as both challenges and opporrtunities. There are already some progress in these directions, e.g., Dopamine, TStarBots, MOREL, GQN, visual reasoning, neural-symbolic learning, UPN, causal InfoGAN, meta-gradient RL, along with many applications as above.

  1. systematic, comparative study of deep RL algorithms
  2. “solve” multi-agent problems
  3. learn from entities, but not just raw inputs
  4. design an optimal representation for RL
  5. AutoRL
  6. develop killer applications for (deep) RL

It is desirable to integrate RL more deeply with AI, with more intelligence in the end-to-end mapping from raw inputs to decisions, to incorporate knowledge, to have common sense, to be more efficient, to be more interpretable, and to avoid obvious mistakes, etc., rather than working as a blackbox.

Deep learning and reinforcement learning, being selected as one of the MIT Technology Review 10 Breakthrough Technologies in 2013 and 2017 respectively, will play their crucial roles in achieving artificial general intelligence. David Silver proposed a conjecture: artificial intelligence = reinforcement learning + deep learning (AI = RL + DL). We will see both deep learning and reinforcement learning prospering in the coming years and beyond. Deep learning is exploding. It is the right time to nurture, educate and lead the market for reinforcement learning.

Deep learning, in this third wave of AI, will have deeper influences, as we have already seen from its many achievements. Reinforcement learning, as a more general learning and decision making paradigm, will deeply influence deep learning, machine learning, and artificial intelligence in general.

The era of reinforcement learning is approaching. Let’s get ready!

Yuxi Li

Written by

Yuxi Li

Guest Editor, MLJ Special Issue,; Co-Chair, ICML19 Workshop,; DeepRL Overview,; PhD UofA

More From Medium

Welcome to a place where words matter. On Medium, smart voices and original ideas take center stage - with no ads in sight. Watch
Follow all the topics you care about, and we’ll deliver the best stories for you to your homepage and inbox. Explore
Get unlimited access to the best stories on Medium — and support writers while you’re at it. Just $5/month. Upgrade