Revolutionary Reinforcement Learning Services

What Is Reinforcement Learning?

As we describe in our book, reinforcement learning (RL) is a sub-discipline of machine learning (ML) that specializes in teaching machines to execute multi-step, strategic decisions.

Traditional ML automates single decisions, but they don’t leverage any context, nor do they operate over sequences. For example, a traditional recommendations algorithm recommends a single set of products and the algorithms are optimized to improve that single recommendation. But this is the wrong objective. You don’t want to optimize for single placements. You should be optimizing for increased engagement or higher profitability per customer, or whatever your business prioritizes.

RL allows you to train your models to do exactly that; optimize decisions over a period of time towards your organization’s unique goals.

RL is Not a Panacea

RL is very good at solving multi-step, sequential decision making problems. For example, YouTube were able to improve video recommendation performance, at the same time as reducing the number of data scientists required to implement such a solution. But industry is awash with “low-hanging fruit” that is best served with simple cloud native or ML solutions.

RL is More Than Games

The vast majority of examples of RL you can find on the internet are based upon OpenAI’s gym. This comes with many pre-baked RL examples, but unfortunately they are focussed towards academia. There are very few examples of industrial problems (except for robotics). But these applications exist. Take a look at our companion site to see many examples of the use of RL in business problems.

RL is Not Artificial Intelligence

Artificial intelligence (AI) is an academic discipline interested in developing algorithms that produce human-like behavior. Although RL lies at the heart of much of this research, it cannot solve all problems for all organizations in an instant. It still requires thorough research and engineering.

How Does RL Help?

Reinforcement learning (RL) helps businesses solve strategic decision making problems and are easily tied towards core business metrics.

RL is often described in terms of the domain. But in our experience developing RL solutions for organizations like Nestle and CMPC, given the right situation, there are a number of generic benefits that only RL can provide:

• Optimizes the right thing: RL algorithms are directly tied to a business metrics via the reward function.
• Uses context: The right decision now may not be the best decision in the future. RL can learn that subsequent actions may be different to those initially taken.
• Learn how, not what: ML typically learns from discrete results; it doesn’t learn how to get there. RL learns how experts achieve a result by learning optimal strategies.
• Strategies, not decisions: ML produces fixed decisions that do not consider future reactions. They are certainly sub-optimal given the high-level goals of the business. RL learns strategies, which encode how to achieve some future state for the organization. The resulting strategies may surprise you!