Engaging Rugby Players Using ML

RFU Business Challenge

I have recently completed a 13-month placement at IBM as part of my “Sandwich Degree” in Maths. Throughout my placement year at IBM, I was constantly looking for opportunities to apply mathematics and statistics to the work that I was doing. An elective session at the Careers Academy event in the IBM client centre gave me an excellent opportunity to do this. The elective session was run by IBMers at the request of the RFU, and was advertised as the “RFU Business Challenge”. As a keen rugby player, I was drawn to this session.

The First Round — Forming a Solution

The RFU’s business challenge is that there is a lack of engagement in rugby and many young players (particularly aged 16–24) are discontinuing their participation in the sport. The session was a competition in which groups of teams had to suggest innovative solutions to help to increase participation and engagement. All teams had just 25 minutes to discuss and construct a solution for the challenge, followed by presentations which were delivered to all other teams and IBMers within GBS (Global Business Services). These presentations were all capped at two minutes.

During the team discussion, I strongly felt a need for a solution which had a combination of technology and active promotion of the sport. For example, taking a technology-based solution to schools, colleges, and universities. With this in mind and a desire to apply maths and statistics to our solution, it was not long into our discussion when I put forward my idea. I proposed a solution which included a Machine Learning model that would input someone’s physical attributes and personal preferences, and output their ideal rugby position. The idea is to take this model into schools and colleges to get students engaged by seeing what their ideal rugby position would be. This concept can be applied to people who have never played, potential returning players and current players.

As a team, we decided to take this idea on. We also added a VR concept to this solution. The idea would be to have a VR experience in the position outputted by the model. We articulated these thoughts and ideas in our presentation. Two days later, we received the great news that we made it through to the second round, where we would be presenting our idea to employees of the RFU and the Client Executive on the RFU account.

Research and Data Collection — From an idea to a proof of concept

In preparation for the second round of this competition, I decided to take ownership of the statistical model element of our proposed solution. My aim was to show the RFU that this solution was feasible. I knew that the first stage in order to present a tangible prototype of the model was to collect the relevant data on rugby players. I created a survey on “Survey Monkey” and posted it into Facebook pages for teams who I have played for or currently play for. The questions in the survey were as follows.

• What position do you play?
• How tall are you in metres?
• How much do you weigh in Kg?
• In an attacking situation, what you opt to sidestep your opponent or run through them?
• Score from 1–10 your preferences for the following:
• Kicking
• Tackling
• Long spin pass
• Catching high kicks

I received just short of 80 responses on this survey which gave me enough data to create the prototype model. I then began to research how to create a predictive model which outputted categorical variables, ie the rugby position. My experience in creating regression models was limited to models which outputted numerical values of Y, however, I needed the outputs to be nominal. I met with one of my statistics lecturers from University who advised me to look into vector generalized linear models for this problem. After researching these types of models, I loaded my dataset into R and wrote the following code:

The “External data” section is the data from the person whose ideal position we are trying to predict. As a test, I made up values for what I considered to be a typical front row player (Prop/Hooker). The output gave me a list of all of the positions followed by probabilities for each:

As you can see above, the probability that this “person” would be a front-row player is 99.6%. Initially, I was not aware that the model would output as above (I was expecting just a rugby position), but I realised that this output is favourable as it delivers insight into the suitability of all of the positions rather than only outputting the highest probability. Also, it would be straight-forward to edit my code such that the highest probability was outputted.

At this point, I was happy that I had created a proof of concept to deliver in the next round.

2nd Round — Delivery to the RFU

A few weeks later, we delivered our presentation to two RFU employees, an IBM graduate and the Client Executive on the RFU account. I explained the predictive model and displayed the code during the presentation. At the end of the presentation, it was clear that the panel were impressed and could see further development of this solution.

We found out shortly after that we were through to the final round which was taking place at Twickenham. We also received some feedback from the panel that they were keen to learn more about engaging new players to the sport rather than returning players. This feedback seemed reasonable as the variables in my model would be hard to determine for individuals who have never played or do not have much knowledge of the sport.

My next task for this project was to think of new variables which would be applicable to everyone.

Final round — Meeting the CEO

In preparation for the final round, I did some research and met with a colleague from the Systems AI team to discuss potential variables to replace the current ones which would make the model more widely applicable. We came up with the idea of using sports day events. For example,

“Rate your preferences (1–10) of the following sports day activities:

• 100m Sprint
• 200m Sprint
• 400m
• High Jump
• …
• Shot Put ”

Sports day events are ideal to use for this as the preference of certain events can be an indicator for the type of rugby position a person would play. It would be likely that someone who enjoyed 100m sprint would be a winger, and a good high jumper may be a full back, etc. Additionally, everyone has experienced sports day so are aware of their preferences for these variables.

We delivered our final presentation at Twickenham to the CEO, CC0 and various other senior members of the organisation with the feedback from the previous round taken into account. The panel were enthusiastic about our solution and we discussed the possibility of expanding the sports day concept into other sports after the presentation. This would create more links from other sports into rugby.

Announcing the Winners!

All three teams who reached the final were asked to step outside while the panel discussed the different solutions. We nervously waited outisde, hoping that we had impressed the panel enough to have won the competition. When we were finally called back into the board room after what felt like a long wait, we were delighted to hear our team announced as the winners! We were also ecstatic when the CEO said that the RFU will be using our solution to encourage more engagement in rugby!

As a prize, we all received a tour of Twickenham Stadium as well as a rugby ball signed by Joe Cokanasiga, Dan Robson and Anthony Watson.