loubbrad / dodge

Transformer model for predicting the outcome of high-elo League of Legends matches.

dodge

./models

Implementations of various transformer based models for League of Legends Match prediction. A model can trained by executing the command python models/run.py combined --db_path data/matches.db --epochs 20. See the ./models/run.py file for a full list of argument options. The SQLite database should have the same structure my opgg-webscraper produces. I have included a small .csv file of euw matches in the correct format, you may run data/build_db.py to build this file into a SQLite file matches.db.

From my testing, the models in their current form are capable of predicting the outcome of a high-elo League of Legends match with ~63% accuracy on a held out test set. Unlike most other attempts I have come across, these models appropriately mask each player's match history, preventing it from using information it would not have access to at test time (unlike this paper for example).

./notebooks/group_match_prediction_paper

An implementation of the paper Group Match Prediction via Neural Networks for predicting the outcome of a match by modelling champion and player-champion 'usefulness' values. Overall I found that the proposed model develops some understanding of what makes a good team composition. The results, however, don't exceed that of other common models applied to the problem. More specifically, my transformer models and models that take advantage of pre-calculated statistics (champion win rate, player experience) give better results by a wide margin.

Installation

You may install the necessary dependencies using conda as follows:

• cd to dodge/ and run the following commands
• conda create --name dodge --file ./config/req.yml
• conda activate dodge

Note that I have encountered some strange behaviour when using the non-CUDA version of PyTorch (issues with dtypes in dataset.py). If you are going to install the dependencies yourself, I recommend making sure the install the CUDA version of PyTorch.