Leaders looking for new ways artificial intelligence (AI) can provide a competitive edge may have found the 2021 America’s Cup Match as exciting for one team’s groundbreaking use of reinforcement learning as for its radical boat designs and close races.
To remain competitive, sailing teams in the America’s Cup contest, like all businesses, must push the boundaries of what is possible. They also face similar constraints, including a steep development curve and a small window of opportunity, meaning teams can pursue only one or two big experiments to up their performance in the sport’s most important competition.
For the 2021 edition of the America’s Cup, reigning champion Emirates Team New Zealand ventured that reinforcement learning, an advanced AI technique, could optimize its design process. The technique delivered, enabling the team to test exponentially more boat designs and achieve a performance advantage that helped it secure its fourth Cup victory.
Pursuing a winning design with AI
See how the engineers, designers, and sailors of Emirates Team New Zealand used AI to create ideal hydrofoils.
Unlike other types of machine learning, reinforcement learning uses algorithms (which often train AI agents or bots) that typically do not rely only on historical data sets, either labeled or unlabeled, to learn to make a prediction or perform a task. They learn as humans often do, through trial and error. In the last few years, the technology has matured in ways that make it highly scalable and able to optimize decision making in complex and dynamic environments.
Besides accelerating and improving design, reinforcement learning is increasingly being incorporated into a broad range of complex applications: recommending products in systems where customer behaviors and preferences change rapidly; time-series forecasting in highly dynamic conditions; solving complex logistics problems that combine packing, routing, and scheduling; and even accelerating clinical trials and impact analysis of economic and health policies on consumers and patients.
We have seen how quickly the technological environment can shift. Only a few years ago, another AI technique, deep learning, vaulted onto the business scene. Today, 30 percent of high-tech and telecom companies and 16 percent of companies in other industries we surveyed have embedded deep-learning capabilities.
Executives who today understand the potential of reinforcement learning will, like Emirates Team New Zealand, be better positioned to find the edge in their industries (see sidebar “Notable examples of reinforcement learning applications”). Understanding the team’s experience can help leaders gauge where and when to use the technology because many organizations will travel a similar path: implementing more traditional technologies first to solve a problem and then applying reinforcement learning to ascend to a previously unattainable tier of performance. Thus, we begin by recounting Emirates Team New Zealand’s journey, after which we offer ideas for where and how businesses should consider applying reinforcement learning.
Emirates Team New Zealand’s journey to a 2021 victory
Emirates Team New Zealand designers were not new to advanced technologies. In 2010, the team had built its state-of-the-art digital simulator to test boat designs without physically building them. This was a key to the team’s 2017 America’s Cup win, but the simulator had limitations. Multiple sailors were needed to operate it optimally, which was a significant logistical challenge given the sailors’ scheduled practices, travel, and competitions. As a result, designers typically iterated on new designs in the absence of simulator performance data and then tested their best ideas in batches when they could carve out large blocks of time with the sailors. Moreover, the sailors’ performance could vary between tests, as human performance often does, making it difficult for designers to know whether a marginal improvement in boat response was due to a design tweak or to variances in human testing.
As Emirates Team New Zealand prepared for the 2021 match, they knew if they could get an AI system to run the simulator, it would free the designers to test more design ideas faster and more consistently than they could with the digital simulator alone. The team was unsure at the outset if the idea was feasible, but as conversations about the technology swirled, team members agreed: the potential payoff was transformative and made trying worthwhile. Using reinforcement learning, experts from Emirates Team New Zealand, McKinsey, and QuantumBlack (a McKinsey company) successfully trained an AI agent to sail the boat in the simulator (see sidebar “Teaching an AI agent to sail” for details on how they did it).
While design rules for the America’s Cup specify most components of the boat, they leave enough freedom for designers to make radical choices on some key elements such as hydrofoils. These wing-like structures attach to the hull and lift the boat above the water, enabling the vessel to reach speeds of over 50 knots (60 miles or 100 kilometers per hour). Hydrofoils can be a significant factor in the race, but race rules allowed teams to build only six full-size hydrofoils in all.
Using the reinforcement learning–trained agent to control the simulator, Emirates Team New Zealand designers could evaluate thousands of hydrofoil design concepts instead of just hundreds in their quest for a winning design. This gave them valuable insight into how a boat might perform on the water before engaging in a costly build and, in the process, would dramatically reduce the design price tag for future races. In addition, as the Emirates Team New Zealand agents’ knowledge of sailing increased over time, the sailors began learning maneuvers from the agents that they had not considered, enabling them to improve their performance for a given design.
Where businesses can use reinforcement learning
The heart of Emirates Team New Zealand’s challenge was to solve a complex business problem in a dynamic environment where the variables change in unpredictable ways, the ideal end state is only loosely defined, and the only way the system could learn about its environment was to interact with it.
That situation is analogous to problems facing retailers, manufacturers, utilities, and companies in many other industries. For example, whereas once retailers could reasonably expect that past consumer behaviors would indicate future preferences, they now operate in a world where consumer purchase patterns and preferences evolve rapidly—all the more so as the COVID-19 pandemic repeatedly redefines life. Manufacturers and consumer-packaged-goods companies are under pressure to build dynamic supply chains that account for climate, political, and societal shifts anywhere in the world at a moment’s notice.
Each of these challenges represents a complex and highly dynamic optimization problem, which, with the right data and feedback loops, is well suited for solving with reinforcement learning.
The appeal of reinforcement learning for problems with many possible actions and paths is that the AI agent does not need to be explicitly programmed. Because it learns from examples and teaches itself through trial and error, it can propose novel and adaptive solutions, oftentimes faster than humans could do so.
How reinforcement learning works
An AI agent learns through trial and error. In simple terms, the agent performs actions within an environment and receives rewards when it takes the “right” actions. It works to find the sequence of actions that maximizes the cumulative rewards it receives. Data scientists and subject-matter experts define the reward function for the agent. This way of learning is just one aspect of reinforcement learning that makes it different from other AI techniques (see Exhibit 1 and “An executive’s guide to AI” for more on the different types of machine learning).
Emirates Team New Zealand, for instance, was able to test multiple designs simultaneously (something the sailors could never do), test tenfold more designs under more conditions than had previously been possible, and gain insight from the AI agent into new ways their sailors could execute on these boat designs on the water.
Broadly speaking, we see reinforcement learning delivering this value across the business, with potential applications in every business domain and industry (Exhibit 2). Some of the near-term applications for reinforcement learning fall into three categories: speeding design and product development, optimizing complex operations, and guiding customer interactions.
Speeding design and product development
Reinforcement learning can improve the development of products, engineering systems, manufacturing plants, oil refineries, telecommunications or utility networks, and other capital projects. Mining companies could, for example, explore a greater range of mine designs than possible with the other AI techniques used today to improve yield. One automotive manufacturer is already exploring how agents trained through reinforcement learning can enable it to test more ideas for regenerative braking in new electric vehicles, so it can optimize the design for noise, vibration, and heat.
Optimizing complex operations
Reinforcement learning’s ability to solve complex problems gives it high potential for optimizing complex operations. Initially, we see three primary applications of reinforcement learning in this area.
First, reinforcement learning can help organizations identify the right actions to take across a value chain as events unfold. A transportation company, for example, can optimize travel routes in real time based on changing traffic, weather, and safety conditions. A food producer can optimize product distribution worldwide amid daily, even hourly, fluctuating demand and exchange rates, varying shipping routes, and more.
It also can help teams manage complex manufacturing processes. For example, it can allow teams to monitor production in real time, simulating different scenarios and updating key parameters to increase production dynamically. Manufacturers that have already used machine learning to minimize product defects can now expand their insights with reinforcement learning to prevent the rare remaining defects that pop up intermittently with seemingly no common root cause.
Finally, reinforcement learning can power autonomous system controllers by, for instance, continuously monitoring and adjusting equipment operating temperatures to ensure optimal performance or running a robotic arm on the manufacturing floor.
Informing the next best action for each customer
When integrated within personalization and recommender systems, reinforcement learning can help organizations understand, identify, and respond to changes in taste in real time, personalizing messages and adapting promotions, offers, and recommendations daily.
Getting to wide-scale adoption
To be sure, implementing reinforcement learning is a challenging technical pursuit. A successful reinforcement learning system today requires, in simple terms, three ingredients:
- A well-designed learning algorithm with a reward function. A reinforcement learning agent learns by trying to maximize the rewards it receives for the actions it takes. A good algorithm with a properly defined reward function enables an agent to make complex decisions—for example, to take an action now that might appear suboptimal in the short term but would pay off handsomely in the long run.
- A learning environment. Oftentimes the learning environment involves a simulator, or digital twin, that replicates the real-world conditions in which the agent will operate and provides a training ground for the agent.
In some cases, however, the learning environment could be a digital platform, such as a product-ordering system, where an AI agent can repeatedly perform the same (or similar) tasks and rapidly receive feedback about the success of its actions.
- Compute power. Training an agent requires substantial compute resources and specialized infrastructure that can scale out thousands of distributed training jobs, which, even when running in parallel, typically require thousands of hours of compute time.
A few years ago, the cost and complexity of building and training these systems was out of reach for all but a few tech leaders. However, significant technological advances to address these hurdles have made reinforcement learning more accessible to more businesses, and continued evolution of the needed tooling is quickly putting the technology within every company’s grasp.
Costs are becoming manageable
The latest iterations in reinforcement learning algorithms, such as soft actor-critic, are dramatically improving training efficiency, substantially driving down compute costs. At the same time, the cost of compute itself has declined significantly. Companies can now access specialized systems in the cloud and pay only for what they use. Also, new tools and strategies enable teams to manage the compute they use. For instance, resource allocation and development tools now available enable teams to identify the least expensive (or most efficient) compute at any given time for a given purpose.
The latest iterations in reinforcement learning algorithms, such as soft actor-critic, are dramatically improving training efficiency, substantially driving down compute costs.
That said, for the technology to be used more widely, compute costs for reinforcement learning tasks will need to decline further. We expect that to happen in the near future for several reasons, including increasing competition among cloud providers.
Cloud-based frameworks address system complexity
Cloud providers have also ramped up efforts to deliver prepackaged, enterprise-ready frameworks that can be deployed in assembly-line fashion and include the necessary tools, protocols, application programming interfaces (APIs), open-source libraries (such as RLlib), and other technologies to eliminate some of the manual coding and integration work. Frameworks can, for example, enable teams to run training jobs across dozens of systems using a single line of code, rather than having to program this capability from scratch. At Emirates Team New Zealand, the development team drew from such frameworks where possible and then focused on the value-added tasks that hadn’t yet been commoditized.
Work remains to be done. There is not yet an out-of-the-box, single framework for delivering reinforcement learning solutions. We anticipate that something like this will be available in a few years from major cloud providers. Efforts under way in this area include Microsoft’s Project Bonsai, Amazon’s SageMaker RL, and Google’s SEED RL.
How leaders can get started with reinforcement learning
The same foundational practices and organizational and cultural changes in which enterprises are already investing for other AI also apply to reinforcement learning. However, given reinforcement learning’s early maturity and its unique requirements and abilities, leaders should keep some strategies top of mind.
Find the right business problem for experimentation
Start by identifying processes where reinforcement learning might free the business to optimize performance in some way, perhaps consulting Exhibit 2 for some ideas. Ideally, select a process where there is already some type of learning environment that can be adapted for training the AI agents.
In our experience, one of the best ways to know if a given process is ready for reinforcement learning is to ask, “What business challenges haven’t we been able to solve with traditional modeling approaches?” Look for areas where teams are conducting AI projects with other methods but haven’t been able to bring them into production because the environment is too dynamic and the models deliver inconsistent results, require too many assumptions and approximations about the data, or cannot handle the full scope of business needs. At Emirates Team New Zealand, for example, testing loops for new boat designs were constantly interrupted by the sailors’ schedules, and there was a high cost to taking the sailors away from other activities.
The right problem should also be one where it isn’t necessary to know why the reinforcement learning system performs the way it does. For now, these systems are not easily explainable, if at all, given the complexity of the neural networks often embedded in them. Reinforcement learning therefore might not be well suited to situations where regulators or operators require transparency.
Factor in compute costs up front
Outlining the reward function to enable an AI agent to learn effectively requires as much art as science, often making it the costliest part of the development process. Subject-matter experts and data scientists need to constantly refine incentives, commonly known as reward hacking, to figure out how to properly calibrate rewards to enable an agent to make complex decisions optimally.
Teams can use first principles to ballpark potential costs, and leaders should understand and discuss the potential cost drivers with their teams up front to help ensure a smoother process and free teams to focus on the work ahead.
Future-proof your simulator
Many manufacturing and operations-focused organizations already use simulation or a digital twin to tune asset performance and utilization. Even in these industries, however, upgrades might be necessary to enable reinforcement learning. Many traditional simulators are designed to run on a small scale, on premise, with only one simulation running at a time, and a person uses a physical interface, such as a joystick, to control it. Such a simulator will need to be re-platformed onto a cloud environment so it can run thousands of simulations in parallel, and it must be updated with an API that enables AI agents to interact with it.
In all cases, whether building or rebuilding digital simulators, organizations should think beyond their existing use cases and make design choices that provide flexibility in supporting more advanced use cases that might not yet be on their radar. Reinforcement learning technology is maturing rapidly, so such planning will enable companies to deploy new reinforcement learning solutions faster than companies that fail to do so.
Double down on humans
Implementations are most successful when leaders recognize that the greatest value comes from using the technology to augment and expand human performance rather than replace it. Any AI initiative relies on domain expertise to help AI teams properly define the use case, determine which data sources to use, ensure the AI predictions and recommendations make sense and can be successfully integrated into their workflows, and guide change management. In reinforcement learning, domain experts must do all this and more, working with data scientists daily to ideate and test different rewards to build an effective reward function and then monitoring the AI agent’s performance after deployment.
Implementations are most successful when leaders recognize that the greatest value comes from using the technology to augment and expand human performance rather than replace it.
Organizations should also consider whether they need a human in the loop to help guide final decisions. At Emirates Team New Zealand, after the AI agents recommended the top designs from the thousands they tested, the sailors then took the helm of the digital simulator once again to test the best hydrofoils and prioritize the final selections.
Identify and manage potential risks
In choosing where to implement reinforcement learning, it’s important to acknowledge employees’ and society’s concerns about the explainability and use of autonomous systems. Our colleagues have written extensively about the unintended consequences that can arise from AI when organizations do not fully understand the possible risks and about the leader’s role in building AI systems responsibly. As reinforcement learning gains traction, leaders will need to build their knowledge around the ethical concerns and interdependencies and how to manage them effectively, so they can guide their company on when to try or not try this new technique.
The technologies that enable reinforcement learning are advancing briskly: compute costs and complexity are declining as the industry evolves toward more adaptive, self-learning algorithms and makes more complex systems available to organizations as high-level services. With this, adoption is increasing, and in a few years, we anticipate that reinforcement learning will become more common in many industries, such as telecom, pharmaceuticals, and advanced industries. Within five years, it will likely be in every leading organization’s AI toolbox, helping companies to uncover innovative strategies and first-in-kind moves that more established techniques may not and to achieve the next level of performance that until now has eluded human reach.