League of Legends Season 11 to Season 13 replay analysis. This repo is split into four main parts:

1. Patch 11.9 and Patch 11.10 single game analysis. This is referred to as TLoL-Prototyping
2. Patch 11.21 multi-early game analysis (Mainly focusing on Miss Fortune). This is referred to as TLoL-Pilot
3. Patch 12.2 multi-early game dataset built for reinforcement learning
4. Patch 13.23 early game Ezreal dataset built for supervised learning

To use the included notebook, you need to decompress the relevant replay using 7-Zip. Be warned, each replay file is around 1GB because of the inefficient way the replay files were generated. Then, load up the *.ipynb notebook file using either Jupyter Notebook, Google Colab or another notebook application and enter the name of the decompressed file.

The included notebook demonstrates how to download one of the datasets from Google Drive and how to then parse the data for reinforcement learning. This will include a basic ML agent in the future.

For a full explanation for how these datasets were generated, refer to the following blog post for very in-depth information.

If you would like to create your own dataset, refer to the following GitHub repo.

This database contains 191 early game surrenders (3.5 minute surrenders) with every single game object within the game recorded 4 times a second (4 observations a second). The actions of each champion can also be inferred as the observations per second is high enough and because the ID's of the spells are contained so they can be associated with players. A much larger dataset will be released in the future which should hopefully allow the development of a human-level League of Legends deep learning agent. Obviously this will favour supervised learning methods but it could also be used as a springboard for offline reinforcement learning techniques as well.

This table shows the top 10 champion occurrences within the dataset.

Compressed Filename: 191-EarlyFF.7z

Decompressed Filename: early_surrenders.db

This database is roughly 455MB (and 6GB uncompressed) and is stored on Google Drive instead of this repository.

NOTE: Roughly half of the size of the database are indexes.

Google Drive Link

These databases contain 987, 728 and 773 games respectively, for a total of 2488 early games. Both databases contain early game replays (first 5 minutes of each game) with every single game object within the game recorded 4 times a second (4 observations a second). Just as above, actions can be inferred by checking object names and net_ids (Network ID, Riot uses this to uniquely identify a game object within a League of Legends game hosted on their server).

Together, this dataset has around 2,985,600 frames for Miss Fortune and several million for many other champions as well. The dataset was curated from a larger dataset by picking the games where the Miss Fortune player lived the longest. This feature had the best correlation with winning (64.4% win rate for this dataset). As the dataset overall has a 64.4% win rate in roughly Diamond II, this ensures the quality of the gameplay is as high as possible without creating a more complicated system to determine the quality of the gameplay.

Compressed Filename(s): 1k-MFLongevity.7z, 750-MFLongevity.7z, 800-MFLongevity.7z

Decompressed Filename(s): For each file in archive => EUW1-{game_id}.db

The databases are roughly 2.04GB, 1.54GB and 1.63GB (25.3GB, 19.2GB and 20.5GB uncompressed, respectively) and is stored on Google Drive instead of this repository.

NOTE: These databases are 7-zip archives with a separate SQLite database for each replay. None of the databases contain indexes.

1k-MFLongevity - Google Drive Link

750-MFLongevity - Google Drive Link

800-MFLongevity - Google Drive Link

This dataset is comprised of 833 early games of Jinx players from Challenger EUW from patch 12.2. Each replay has undergone a complex procedure to transform it into a format which is suitable for bulk analysis or directly training a reinforcement learning agent to play League of Legends (or at least for the first 5 minutes).

In total this dataset contains 1,189,786 frames, which is comparable to ImageNet. This should be enough to at least experiment with basic Imitation Learning algorithms such as Behavioural Cloning, or perhaps other algorithms such as Generative Adversarial Imitation Learning (GAIL).

The uncompressed archive is 889.8MB and the uncompressed folder is 7.7GB. This should be a manageable dataset for most researchers to experiment with. All of the features in this dataset have been saved as float16 instead of float32 to halve the filesize. Fortunately, this also has virtually no impact on accuracy as every single variable in League of Legends doesn't need more than 16-bit floating point representation.

Compressed Filename(s): jinx_833_ml_db.zip

Decompressed Filename(s): full_db/, For Each file in archive => EUW1-{game_id}_jinx_100.pkl

833-JinxML - Google Drive Link

833-JinxML Raw - Google Drive Link

The table below contains useful metadata for this dataset:

This dataset is comprised of nearly 10,000 Ezreal early games. This will be used as the pilot study for training an agent to play Ezreal using supervised learning. There are roughly 5.4 million frames in this dataset which should be adequate to build a simple bot with very basic performance. Especially when using the methods from the JueWu-SL paper.

Raw DB's Google Drive.
Converted into Observations / Actions Google Drive

Enter EUW1-5270795542.rofl.json after decompressing the replay file into the notebook to analyse this replay.

Compressed Filename: EUW1-5270795542.rofl.7z

Decompressed Filename: EUW1-5270795542.rofl.json

Enter EUW1-5237530168.rofl.json after decompressing the replay file into the notebook to analyse this replay.

Compressed Filename: EUW1-5237530168.rofl.json.7z

Decompressed Filename: EUW1-5237530168.rofl.json