IBM Watson has been used to assist some of the world’s biggest companies and help tackle some of humanities greatest threats. Now, it’s being used to assist Leatherhead Football Club, a team that play in the 7th tier of English Football and are made of up delivery drivers, car salesmen and shop assistants. They’re certainly not people you’d associate with the latest Artificial Intelligence technology, however, since the start of the 2018/19 season, that’s exactly what they’ve used, and it worked.
IBM’s Emerging Technology team, part of IBM Research, have been working in partnership with Leatherhead’s management to develop tooling that can assist their pre/post-match analysis and opposition scouting. The platform is designed to give the coaches and players easy access to data, including video footage, Opta data and valuable insights and analytics.
Low Barrier of Entry
A key challenge when working with any new technology is its complexity. Nikki Bull, Leatherhead manager, flagged at the start of the season, that anything that IBM provided needed to be easy to use. Nikki and his team wouldn’t have the time to go through training on how to get to grips with the technology and certainly wouldn’t have the technical know-how or desire to understand the ins-and-outs of artificial intelligence.
Our solution? Artificial Intelligence.
Ask a Question, Any Question.
To use the tooling the team simply type their question or query. Under the covers, a trained instance of IBM Watson Assistant will read it, understand it and present the relevant data and results on screen.
When Leatherhead want to review the shots that they had in their game against Bognor Regis, they simply type “Show our shots against Bognor”.
The playlist that the tool returns when asking Watson to show the shots that Leatherhead had in a particular match.
This returns a playlist that includes the relevant video footage of each shot, it also provides heatmaps to show where those plays have taken place across the pitch (attacking play goes from left-to-right) and visualisations showing when during the game the shots took place, the duration of the plays and the distribution of pass lengths.
Watson can understand a variety of entities including the team and player names, all of the in-game events, areas of the pitch, times, formations and scores. The results, analysis and data are then presented in a clear interface to the user, allowing them to review the plays and conduct their own review with Watson’s assistance.
Demonstration of the football terminology that Watson can interpret. In this case, location on the pitch, the score and formation. This view also shows the play heatmap (where on the pitch the returned plays have taken place) and a breakdown of the play types that have been categorised.
Context is Key
When analysing a game of football, or any sport for that matter, context is key. It’s great to understand that a particular player scores most of their goals from the right-hand side of the box at about 10 yards out – but how did the ball get there in the first place? Did it come down the right wing? Left wing? Did the team kick the ball long, or play patient build-up play? Do the goals tend to come when playing a particular formation, or later in the second half?
Review of the plays that Leatherhead have done containing a long ball. Breakdowns include when in the game they’ve taken place, the duration of the plays and a heatmap showing on the pitch where the most activity has taken place. The tool also shows that just under 6% of plays in which Leatherhead perform a long ball result in a shot on goal.
Picking out these patterns of play is key to beating an opposition. If you can understand their strengths and weaknesses, the methods by which they plan on attacking or how they generally concede their goals, then you stand the best chance of creating the best tactics to raise your game and get the win.
When asking Watson for the shots that a particular team has, it doesn’t just limit itself to the actual shot itself, it provides context through the events that lead to the shot. The context is gathered through a mixture of rule-based models and machine-learning models that decide when a play starts and ends.
Modelling the Flow of Football
We use a rule-based model for static starts (e.g. penalties, free-kicks, corners) and static ends (e.g. the ball going out of play, a goal). The machine-learning model then comes into play to decide whether events between, such as interceptions or blocks would constitute the end/start of a play.
Given knowledge of the full play, we can enquire as to which players were involved throughout the entire play. The right-side here shows when the players are involved, additional data that Watson can use in categorising playing styles for a particular player.
The machine learning model is trained with historical footage and Opta stats. Annotated datasets were produced that detail when analysts consider a play to start and finish. In support of this data, information on the previous events, next events, play length, position on the pitch and much more were accumulated to make comprehensive training data.
With the start and end of each play in a match defined, we then have another model that categorises the high-level label associated to that play. These labels make analysing the data easier to interpret and given that Watson can understand the terminology used, also provides the coaches the ability to review particular scenarios/categories with ease.
Given the context of a play, the tooling can then present valuable information to the analysts and coaches based on the questions they’ve asked of Watson. In turn, this can then be used to decipher the patterns of play that Leatherhead should undertake in order to score more and concede less.
Getting ahead of the Opposition
In the top leagues, teams have their own dedicated opposition scouts. They will attend matches, review footage and provide scout reports to the management so that the team can plan accordingly. In the world of semi-professional football, this privilege isn’t always available, and if it is, often comes at a financial cost that teams struggle to afford at scale. Most semi-professional teams go into their games and know relatively little about their opponent.
Example of the commentary available through social feeds for a game.
Watson has the ability to digest, and importantly understand, huge amounts of written text. Published match reports and social media traffic provide plenty of detail on how a particular opposition team plays. When we point IBM Watson to these feeds, it can review entire seasons worth of matches for Leatherhead’s upcoming opposition and produce a clear summary of the findings. Watson can’t do this out of the box though. We first need to train Watson to understand complex football terminology and we can do this through IBM Watson Knowledge Studio (WKS).
Training Watson in Football Punditry
IBM Watson Knowledge Studio provides developers the ability to train Watson on how to interpret and read particular nomenclature and terminology. It’s possible to do this using rule-based models, defined through RegEx, or, you can train Watson through a machine learning model, such that Watson can take in context of the annotations. In our case, we use a bit of both.
We begin by building a model of the entities that we want Watson to interpret, these can then be labelled into particular entity “types”, in our model, these include “football_formation”, “football_score” and “football_player”. Once defined, we can then add more clarity with entity “subtypes”, for example, “football_event_card” would breakdown into the subtypes “yellow” and “red”. Then, we being the training process through “annotation tasks”. This involves teaching Watson, given various examples, which pieces of text are the player names, team names, references to formations, etc.
IBM Watson Knowledge Studio – Here, the developer is teaching Watson where instances of the “football_timestamp”, “football_player”, “football_tactic” and “football_team” have occurred in this particular tweet.
Once we had enough examples, we published the model, meaning we can use it in IBM Watson Discovery. The WKS model provides Discovery the ability to interpret the football-specific terminology. Now that Watson had an understanding of football terminology, we could upload the relevant documents and feeds for Leatherhead’s upcoming opposition. In each case Discovery then allows us to query the data, looking for patterns, relationships and correlations in the findings, meaning that Watson could begin to automatically create opposition scouting reports.
The analysis in IBM Watson Discovery provided the management with a comprehensive view of their opponent’s recent games. Example details included the most influential players, tactics for attacking and the balance between the left/right wing with the manager, Nikki Bull, then being able to adapt his tactics accordingly.
The scouting reports and the Opta analysis through Watson Assistant were designed to work together. In one particular case, Leatherhead were due to play Hitchin Town in the FA Cup 4th Qualifying Round. The analysis through Watson Assistant had flagged a weakness in Leatherhead’s left-back position, this coincided with Watson Discovery showing that Hitchin have a tendency to go down very easily in the penalty area. Unfortunately for Leatherhead, these insights would come true in the game when their left-back gave away a penalty, resulting in Leatherhead losing the game and exiting the FA Cup.
Assessment of the season
Our key objective at the start of the season was to demonstrate how IBM’s technology, in particular IBM Watson, can be used to augment the intelligence and expertise of Leatherhead’s coaches and players.
The key challenge was going to be how to introduce such complex technology to users who are not tech-savvy. At the halfway point through the season, the tool is used weekly by the management, used as evidence in training sessions to help the players develop and even used by the players directly themselves, something that is just not seen in modern football analysis. It’s proved hugely valuable to the team, at both the management and player level.
At the point of writing this article, ‘The Tanners’ have just won their final game of the season, a comfortable 4-0 win against big-spending rivals Kingstonian, ending their 2018/19 season on a high and leaving the team sat 8th in the league, a great achievement having been in the relegation zone for the beginning of the season.
Over the season, Leatherhead have lost key players to higher-paying rivals, key injuries haven’t helped either, however, IBM Watson, working alongside Bull and Martin McCarthy, Assistant Manager, have ensured that Leatherhead continue to raise their game. With Watson’s assistance, Leatherhead ascended 12 spots up the league, rounding off an incredibly successful 2018/19 Season.
How we made a game of learning Qubits . In this article we will briefly show the tool we built while learning quantum computing; alas, this article is not about teaching these concepts. We will be covering these in a series of articles soon. At the start of our journey we were learning […]
In March 2019, Dan Cunnington from Emerging Technology was deployed to Puerto Rico to help facilitate the field testing of the winning Call for Code solution – Project Owl. As part of the grand prize, Project Owl won a field testing opportunity supported by IBM Corporate Service Corps to field test their winning solution on […]
Introduction Disabled or disadvantaged passengers have every right to fear travelling with multiple high profile cases of unfair and appalling levels of service evident in the media, particularly for rail passengers. Our team have been tackling this problem recently in collaboration with the IBM Travel and Transport team and 50% funding through Innovate UK’s call […]