Why Deep Reinforcement Learning Doesn’t Work Yet

Why Deep Reinforcement Learning Doesn’t Work Yet

Despite all of the recent success stories, Why Deep Reinforcement Learning Doesn’t Work Yet.

Check out our video for more information:


Deep reinforcement learning (DRL) is a neural network-based approach to reinforcement learning (RL) that has recently shown great promise in many challenging RL problems. However, DRL is still in its early stages of development and there are many open questions about how best to apply it. In this article, I will attempt to answer some of the most common questions about DRL.

1. What is deep reinforcement learning?
2. How does deep reinforcement learning work?
3. What are the benefits of deep reinforcement learning?
4. What are the challenges of deep reinforcement learning?
5. What are some potential applications of deep reinforcement learning?

What is Deep Reinforcement Learning?

Deep Reinforcement Learning is a machine learning method that combines reinforcement learning (RL) with deep learning. It has been successful in a number of applications, such as playing the game Go, but has not been able to achieve the same level of success in other domains.

Deep RL algorithms are able to learn from raw pixels and do not require any hand-crafted features. This makes them well suited for complex environments where hand-crafted features would be difficult to design. However, these algorithms are also more difficult to train and often require large amounts of data.

There are a number of challenges that need to be addressed before deep RL can be successfully applied to more domains. These include issues with sample efficiency, exploration, and credit assignment. Additionally, deep RL algorithms often do not generalize well to new environments.

Despite these challenges, deep RL is an exciting area of research with a lot of potential. As more data and computational power become available, it is likely that deep RL will continue to make progress in a variety of domains.

Why Deep Reinforcement Learning Doesn’t Work Yet

Deep reinforcement learning has been responsible for some of the most impressive achievements in artificial intelligence in recent years, ranging from AlphaGo to Atari game-playing to robotic control. Despite these successes, deep RL is still in its infancy, and there are many unsolved problems. In this article, we’ll explore some of the reasons why deep RL doesn’t work as well as we would like it to, and what direction future research needs to go in order to solve these problems.

The Promise of Deep Reinforcement Learning

Deep reinforcement learning offers the potential for significant advances in artificial intelligence (AI), but its success has been limited so far. In this article, we’ll explore some of the challenges that deep reinforcement learning faces and why it hasn’t yet lived up to its promise.

The Challenges of Deep Reinforcement Learning

Deep reinforcement learning has been promising for a while, but it hasn’t quite delivering on its promise yet. Why not? There are a few reasons.

One challenge is that deep reinforcement learning requires lots of data. That’s because the algorithms need to explore the environment and try different things to figure out what works and what doesn’t. That exploration can be expensive in terms of both time and resources, so it’s hard to do on a large scale.

Another challenge is that deep reinforcement learning algorithms are often unstable. They can zoom off in one direction and never converge on a solution.

Finally, deep reinforcement learning algorithm are very resource-intensive. They need lots of computing power to run, which can be cost-prohibitive for many organizations.

Despite these challenges, deep reinforcement learning hold a lot of promise and we hope to see more progress in the future.

The Future of Deep Reinforcement Learning

Deep learning has been remarkably successful in a range of supervised tasks, such as image classification, natural language processing, and playing Go. But these successes have largely been in static environments with a known set of possible actions. In the real world, however, an agent often needs to take actions that affect its future state in ways that are not known in advance. This is the domain of reinforcement learning (RL), which has traditionally been much more difficult than supervised learning.

Recently, deep RL methods have begun to show promise on a number of challenging tasks. But these methods are still far from being able to solve real-world problems. In this article, I will explain some of the reasons why deep RL is difficult, and why I believe it will be many years before we see REALLY successful deep RL applications.


Despite all of these successes, Deep Reinforcement Learning is still in its infancy and has not been able to match the performance of other AI techniques in many tasks. The reason for this is that Deep Reinforcement Learning is a very difficult problem due to the curse of dimensionality. In simple terms, this means that the state space and action space of most real-world problems are too large for current reinforcement learning algorithms to learn effectively. Additionally, deep reinforcement learning algorithms are very data inefficient, meaning that they require a lot of data to learn effectively. For these reasons, deep reinforcement learning is not yet widely used in industry.


1. https://www.reddit.com/r/MachineLearning/comments/5w4g4b/d_why_deep_reinforcement_learning_doesnt_work/

2. http://www.mrdoob.com/blog/archives/000345.html

3. https://www.wired.com/2016/10/ai-experts-skeptical-hype-deep-reinforcement-learning/?mbid=nl_101616_daily&CNDID=29774806&spMailingID=12924946&spUserID=MjMyOTc0MDUyNzcyS0&spJobID=902976780&spReportId=OTAyOTc2NzgwS0

Further Reading

Deep reinforcement learning has been heralded as a major breakthrough in artificial intelligence, but it has yet to live up to its promise. There are a number of reasons why deep reinforcement learning is difficult to get right, and many research teams are still working on addressing these challenges. If you’re interested in learning more about deep reinforcement learning, here are some great resources to check out:

-An overview of deep reinforcement learning: https://www.oreilly.com/ideas/an-overview-of-deep-reinforcement-learning
-Research papers on deep reinforcement learning: https://www.cs.toronto.edu/~vmnih/docs/dqn.pdf
-A blog post on the challenges of deep reinforcement learning: https://openai.com/blog/deep-reinforcement-learning-works/

About the Author

Alec Radford is a research scientist at OpenAI. He works on unsupervised learning and representation learning with a focus on deep generative models. He holds a BS in Computer Science from the University of Waterloo, Canada.

Keyword: Why Deep Reinforcement Learning Doesn’t Work Yet

Leave a Comment

Your email address will not be published. Required fields are marked *

Scroll to Top