Soccer is often misinterpreted as a game of pure finesse and agility – thanks in part to its most dominant players. Watch Lionel Messi in action, for instance, and it’s apparent that his footwork and situational awareness give him a major competitive advantage. In reality, though, athleticism plays a relatively limited role in whether a team wins or loses. The rest of a winning strategy occurs on a macro level, and involves everything from player formations to substitution timings, lineup synergies, and so on.
This is why, when we ask an artificial intelligence (AI) what makes a successful team, we need to throw out our preconceived notions about individual skill – even if we’re talking about Messi himself.Whenever we try to solve a new problem, our scientific process, by definition, looks to either support or reject a hypothesis. But as humans, with innate biases, we are fundamentally limited in our ability to interpret data and uncover new tactics. In fact, while you and I might view Lionel Messi as the perfect soccer specimen to study, an unbiased AI might even toss out his approach altogether.
Recently, a team of CalTech scientists ran into the problem as they worked to interpret player data in an attempt to synthesize new strategies. To get around innate biases, they use a technique called ‘ghosting’ – which, for the CalTech team, allowed their AI to interpret a large set of data without supervision:
Though it doesn’t have a concept of a left-back, the AI can notice that about 1/11th of the tracker points play very similarly to one another, and that they tend to conform to reliably different rules than data points that seem to be in other, equally reliable categories. These automatic position assignments saw through the variability of different individuals’ approach to their positions, as well as any transient changes in play, such as when a goal is near and everybody is playing defense. What this means is that they can have their AI quickly learn the strategies of a new team simply by giving it a mess of tracking data.
When an AI interprets a data set so massive, it can become impossible for humans to retrace its steps and understand how it reached its conclusions. This is something that scientists are working to fix, as they attempt to find ways for AI to ‘explain itself’ to be comprehensible for humans. But just because we may not be able to understand the entirety of an AI’s decisions doesn’t mean it isn’t providing useful insights. We saw this with Google’s board-game-playing AlphaGo, which triumphed over one of the world’s top Go players using strategies and patterns that had never been seen before.
Solve industry problems
The effects of machine learning in games give us just a peek at its paradigm-altering power. Training AI on the super computer that is the cloud, with exposure to vast amounts of data, we can bring it to life in business applications seamlessly and solve very specific industry problems in a powerful and unique way. Apply these methods to retail, and we can give customers the products they want before they want them. Apply them to healthcare, and we can make hospitals run with optimal efficiency – allowing medical teams to make the most of their time and provide better care.
Society is changing, one learning algorithm at a time. By striking a balance between unsupervised machine learning and supervised learning – informed by our understanding of the industries we serve – applications in the age of networked intelligence offer both the macro- and micro-level insights to help organizations become dominant players in their own respective arenas.
Image credit - Robot kicking football on pitch with mountain behind © Kovalenko I - Fotolia.com