Queensland University of Technology (QUT) researchers have developed a machine learning system that can predict where three top male tennis players will hit the next ball at any point in a match.
Their model analysed thousands of shots made by Novak Djokovic, Rafael Nadal and Roger Federer at the Australian Open, before it is able to “figure out a player’s style” said QUT’s Speech, Audio, Image and Video Technology Laboratory researcher Dr Simon Denman.
The machine learning system – called a Semi Supervised Generative Adversarial Network architecture –was trained on Hawk-Eye data from the 2012 Australian Tennis Open, provided by Tennis Australia.
More than 3400 shots for Djokovic, 3500 shots for Nadal and 1900 shots by Federer,were fed into the system, with the researchers tagging them for whether the shot was a return, a winner or an error.
“After about 1000 shots, the model has a pretty good idea of what is going on,” Denham said, “Once it’s got three matches it’s pretty solid.”
The model – which can predict about 2000 shots per minute – also takes into account when in the match a point is being played. This is necessary as a player’s shot selection will be drastically different depending on whether they are opening the first set or clawing back points in game five.
“We train the model in order so that it sees the shot from first round, to the second round and third round – so it builds on experiences like a human does,” Denman said.
“We are trying to mimic what we think what the tennis player’s brain might be doing,” he added.
In order to give the model ‘player brain’ like capabilities, two forms of memory were utilised: episodic memory and semantic memory.
“Episodic memory is effectively individual recollections. It’s being able to go back and recall each individual stroke and what happened. Semantic memory is much more abstract. It’s the overall learnings that came about from many, many, many instances of the episodic memory,” Denman explained.
“Then those two memories work together given an input stimulus. They each pull out something relevant from their own memories and use that to help reinforce the prediction of what’s going to happen. The episodic memory can look at the input and say `I’ve seen shots like that here, here and here – here is something useful’. The semantic memory says `we should hit it over to that part of the court because that’s a good tactic’,” he said. “That then helps to guide the output generation.”
Some players were easier than others to predict. The toughest to nail was Federer.
“Who is perhaps the most versatile. It struggled the most to predict him. He can do anything, so the model was more often wrong about him. Given how hard Federer’s game is to predict, it just adds to the credit of someone like Stefanos Tsitsipas who managed a victory against Federer in the Australian Open on the weekend,” Denman said.
The research team, which includes PhD student Tharindu Fernando, Professor Sridha Sridharan and Professor Clinton Fookes, all from the Vision and Signal Processing Discipline at QUT, believe the model could be applied to top level coaching regimes within the decade.
It could also be applied to virtual reality, allowing users to go head to head with accurate simulations of the world’s best players.
Potentially other sports could be analysed in the same way, Denman added.
“This example here is a single trajectory with the ball, but there’s no reason why these techniques could not be applied to team sports such as soccer where you are tracking every singer player from both sides,” he said.
“Sport is good in that sense for any sort of machine learning problem because you’ve got all of these constraints and rules as to what can happen, which can simplify some problems compared to other domains.”