Using Machine Learning to Predict Current Ability

Machine Learning is a subset of Artificial Intelligence and the fundamental idea behind it is to train a model by learning patterns from data. We can then use the model to make predictions. So how does this relate to Football Manager and CA?

Well, if we can get attribute and position ratings for players as well as their CA, we can train a model to uncover the underlying patterns that map the inputs (attributes and position ratings) to a particular output (CA).

Once the model is trained, we can provide it with some unseen test data (a player's attributes and position ratings) and it will predict the CA.

Regression

This type of ML problem is called regression. Regression is a type of supervised learning where the goal is to predict a continuous numerical value based on input data.

In other words, regression algorithms are used to model the relationship between input variables (also known as features) and a target output variable.

The main objective is to find a function that best captures the underlying patterns in the data, allowing the model to make accurate predictions for new, unseen data points.

A simple example may be a model to predict someone’s age based on their height. We would train the model by collecting training data by getting a group of people, measuring their height and asking them their age.

The ML regression algorithm would then come up with a relationship between the feature (height) and the target (age).



Once we had trained the model, we could then find a new person (unseen data), measure their height and use the model to predict their age.

Of course, it may not be 100% accurate, but a well-trained model would do a good job of predicting their age.

You can imagine that training the model with more data will enable it to make more accurate predictions (consider the training data collected from 5 young children compared to training data from 1000 people including one-week old babies, toddlers, teenagers and fully-grown adults).

Player Data

This idea can be extended to predicting the CA of players. In the previous example we had one feature (height) and one target (age). In the context of our problem, we have many features (each of the player attributes and position ratings) and one target (current ability).

So we train the model by providing a sample of players (their attributes, position ratings and CA).



Once trained, we can then use the model to predict the CA of a player by providing the attributes and position ratings.



For the purposes of training the model, we extracted the relevant attributes and current abilities of around 25,000 outfield players from the game.

By relevant attributes, we mean the attributes that are known to contribute to CA:

• Technical - Corners, Crossing, Dribbling, Finishing, First Touch, Free Kicks, Heading, Long Shots, Long Throws, Marking, Passing, Penalties, Tackling, Technique
• Mental - Anticipation, Bravery, Composure, Concentration, Decisions, Leadership, Off The Ball, Positioning, Teamwork, Vision, Work Rate
• Physical - Acceleration, Agility, Balance, Jumping, Pace, Stamina, Strength
• Weaker Foot

As well as the position ratings (1-20)

• DL, DC, DR, WBL, WBR, DM, ML, MC, MR, AML, AMC, AMR, SC

A histogram of the CA distribution of the players is shown below. Note the very few players with high values of CA. This has implications later for predicting the CA of top players; they are essentially outliers to the model so there is not much data for the model to be trained on.

Training the Model

For the purposes of training the model, the player data was split into two groups, training data and unseen testing data. 75% of the players were used for training and 25% of the players were used for testing the trained model (i.e. their predicted CA was compared to the actual target CA).

Following training and tuning the model parameters, an accuracy of 98% was obtained.

A plot showing the target CA against the predicted CA is shown below.



Each blue circle represents a player and the red line represents the situation in which the predicted CA is exactly equal to the target CA. In an ideal world, all the blue circles would lie on that red line.

Note the small group of players at 175+. Despite the fact there are only a few of them, the model still accurately predicts their CA.