Reinforcement Learning Platform for Enterprise Growth

Vrajesh Bhavsar

Nov 28, 2018

If you follow technical updates on machine learning or artificial intelligence, expect to be overwhelmed over the next few weeks. NeurIPS 2018 begins next week! Based on my experience last year, the knowledge bottled among the academic community will become accessible to developers and entrepreneurs to further adopt machine learning in technologies we use every day.

There is no doubt that enterprises are catching on to the value of machine learning as well, because of its impact on their core business. In fact, in McKinsey’s report about AI being the next digital frontier, they surveyed 3,000 AI-aware C-level executives, across 10 countries and 14 sectors and found that 20 percent currently use AI-related technology at scale or in a core part of their businesses, with nearly 40% contemplating to use similar technologies in the coming future.

Similar trends are also supported by Gartner, with AI-driven development being one of the key trends among Gartner’s Top 10 Strategic Technology Trends for 2019. In fact, by 2021, Gartner projects that 40% of new enterprise applications implemented by service providers will include AI technologies.

As an enterprise app developer or decision maker, if you are considering AI adoption, this blog will summarize why you should specifically consider reinforcement learning (RL) and the challenges in doing this on your own.

Why RL?

As you consider the details of what AI technologies you need, you will come across different types of learning:

In summary, supervision related learning concepts rely on learning from other similar examples from the past. Typically, this is available as labeled or structured data from controlled datasets. While in RL, the learning agent must interact with the environment to discover the sequence of actions that yield maximum rewards. Here, an agent doesn’t expect labeled data directly, but instead, relies on discovery because of a feedback loop with the environment in taking different actions to maximize reward.

A fundamental distinction between the three supervision-related learning methods and reinforcement learning is that in reinforcement learning, the future data that you collect from your users is heavily influenced by the decisions that you make with your machine learning models. One obvious downside of this is that you have to manage how the models you learn now will influence what models you can learn later; the upside is that your future data changes because your models are continuously working and planning to deliver better results to your end users.

bFzQQPSJRezRRFN7oDpXX5FaBSNHfKLW9Gi5mBVXDaikEJCNf0G2_dn3Fj6J7n1iNHajmFzQWSGhtcECdmjO8ejd3xQg5UBtJDaGcqPtNEEYG7xlQlsIFXxRQ04As71_XmP7tnoP

Because of this ability to work with different environments, researchers have explored applications of RL for many domains - from chatbots to mobile gaming to robotics! OpenAI, one of the non-profit AI companies, recently shared details on their RL platform and how it can be applied to gaming simulations. There is also an active research community that believes in applying RL towards artificial general intelligence (AGI) since it mimics human methods of learning. Richard Sutton & Andrew Barto discuss this in their book on Reinforcement Learning.

RL and non-RL learning systems are useful for different types of problems. Since enterprises are driven by metrics & growth, we believe a large set of their use cases are well suited to be optimized by RL. Here are some other factors to consider:

Companies with deep ML expertise are investing in RL infrastructure of their own. Both Google and Facebook have shared research about how their RL platforms are solving a variety of use cases internally. But this leaves an open question for enterprises that are unable to tap into such research - should they care about the innovations in RL at this time? Should they be developing their own RL platforms and exploring its impact on their production applications?

Why developing your own RL platform is hard

While research unlocks new ideas, it takes some time and talent to put such ideas to work at scale. If you consider the same for the concepts of reinforcement learning, there are aspects of the algorithms themselves that are hard to implement on their own. Even if you are prepared to tackle this hard/open problem, there are other challenges to consider:

In a related post, our systems engineering lead goes into the details of these challenges and how amp.ai has solved them. So even if you bring all ML experts, big data experts, and systems experts together, the end result is going to be quite time consuming, expensive and niche for your company. Meanwhile, the research keeps moving forward and you have got to keep pace with it as well!

The other big consideration in rolling out your own RL platform is to have some great insights about what problems you are going to solve with it. In case you’ve ever wondered, not all enterprise problems are about gaming simulations or virtual robotic worlds! Some of the enterprise teams we meet are not yet sure whether their use cases are right for applying RL. At Scaled Inference, we have been guiding these discussions for different use cases and have come up with design criteria that are helpful to get you started. Besides the use cases in the experience optimization space, we have also been discussing the use cases in infrastructure and IoT domains. Watch this space for more blogs on these topics.

Enter amp.ai

At Scaled Inference, we have developed a reinforcement learning based platform called amp.ai. We believe it is world’s first production-ready SaaS platform, that is well packaged and broadly applicable, available to all enterprises today!

At the core of it, we have a generic agent that can bring RL capabilities to your enterprise apps right out of the box! You don’t have to worry about how optimization is done or what infrastructure capabilities you need. With simple API interfaces, you can start integrating RL in your use cases. If you do want to review the details, you always visibility into what the system is doing from our dashboard. All this results in a huge time-to-market advantage.

You can learn a lot about the platform itself and how to set up a project yourself via some of our other blogs. I highly recommend you review the blog that provides a technical overview. We have also written about how you can use the platform to maximize e-commerce revenue growth or how to easily implement personalized offers & prices. Happy reading!

Conclusion

Applied machine learning is going to be a large part of digital growth in the years to come. Many enterprise use cases can benefit from RL today if they are designed and deployed on the right platform. Scaled Inference has developed a reinforcement learning based SaaS platform that is ready to take on enterprise workloads and deliver results. More teams are discovering the value of working on a production-ready platform and experience faster benefits of RL on their use cases.

If you also share the same vision of making AI useful & accessible to everyone, you are welcome to join us! Contact us to get a demo of this industry-leading enterprise RL platform and ping me if you’ll be at NeurIPS next week!