As football fans, it’s easy to watch our favorite teams play and be couch coaches and recliner refs. We’re able to watch the action on the field and make sense of the chaotic movements of the players and figure out the design and intention of the plays. Even for the casual fan, their brain is able to make sense of the basic strategy and rules of the game. But, how did we get to this point? How did we initially learn what is actually a very complex sport? For those that actually play the game, that learning and reaction process is critical to on-field success.
Now, researchers at Oregon State University are trying to use artificial intelligence to get at that learning process from a machine’s point of view. The design of their experiment was ambitious; have computers watch the OSU Beavers football team on video, learn the game, then be able to design plays and manipulate players in a football simulation.
This new system would need to combine computer vision, machine learning and automated planning as sub-components to the learning process. “This is one of the first attempts to put several systems together and let a computer see something in the visual world, study it and then learn how to control it,” said Alan Fern, an associate professor of computer science at OSU.
Fern’s team not only hopes to someday help the Beavers coaching staff, but is most interested in the generic learning processes and how machine-enabled artificial intelligence can observe complex processes in manufacturing, business and logistics and then make them more efficient.
“Football actually makes a pretty good test bed, because it’s much more complicated than you might think both visually and strategically, but also takes place in a structured setting,” he said. “This makes it quite analogous to other potential applications.”
The research was published in a recent issue of AI Magazine.
The two key parts of a knowledge transfer system are the source and the target. The source is the environment that a human or machine agent has available to receive input, while the target is the new output generated as a result of the learning process. The researchers cite the progress made in classifying transfer systems by psychology researchers.
First, there is a distinction between reproductive and productive systems, where source data is reproduced for a similar task versus the production of a new result. Second, systems can vary by whether they transfer knowledge itself (either declarative or procedural knowledge), or a problem-solving process. Third, the amount of processing done by the system to either knowledge or a problem-solving process to get it ready for the target system can vary. As the authors write, “High-road transfer requires the system to actively retrieve, map, or otherwise infer relationships and similarities among knowledge components in the two tasks. Low-road transfer is automatic in the sense that the system need not apply any computational effort to apply the transferred knowledge in the target.”
For their football play learning process, they categorize their new system as a productive knowledge-based transfer with element of both high-road and low-road processing. The initial research focused on twenty passing plays which is just a small sampling of the entire game of football. The early results have been encouraging in terms of a target system that can recommend a play based on game circumstances.
“Using football, we created learning algorithms that allow the computer to see the plays, analyze them and learn from them,” Fern said. “Ultimately these systems should be able to see what is happening, understand it and maybe even improve upon it.”
So, while this research may help design a more efficient workflow or process over time, it also emphasizes the wide gap that remains in understanding how the brain of even a high school quarterback can learn to read defenses and adjust on the fly. These researchers acknowledged the complexity of the game of football with the large number of variables to be evaluated prior to a decision.
The role of technology as a decision support tool is being considered in several different fields. In medicine, electronic medical record systems help physicians with care reminders, drug allergies and evidence-based diagnostic suggestions. In air travel, pilots are assisted by numerous monitors and reminder systems. In both cases, both doctors and pilots are in charge but have accepted AI technology as a tool to help them perform their job better.
What can human players and coaches gain from this type of artificial intelligence work? First, teaching a machine to understand and reason so that it benefits humans is the first task of AI, but second, the knowledge architectures used for these tools might provide some insight on how human intelligence is organized.
For most players and coaches, any opportunity to get an edge over their opponent is at least worth a look. As these training tools evolve, they will become more mainstream just as physical training is required today.
Dr. Fern certainly hopes he can help. “The study of these football plays is helping us to create intelligent computer systems,” he said. “When this is more fully developed, we should be able to actually apply it to football, maybe help coaches analyze an upcoming opponent, let the computer determine what they are doing and suggest a strategic nugget to the coach.”