cmnemoi / OWREstimator

An app allowing you to predict the best possible time possible for a game speedrun.

TAS Predictor : Project Overview

An app allowing you to predict the best possible time for a game speedrun.

• Created a web app that estimates the best possible time of a speedrun (MAE ~ 388 seconds).
• Extracted 3800+ game runs from Speedrun.com using its API and Python.
• Scrapped 2000+ game runs from TASvideos using Beautiful Soup and Python.
• Engineering features from the time of the world records, number of runners and released year for each game put on Python, Libre Office Calc/Excel and Streamlit.
• Optimized linear, lasso, ridge, random forest, gradient boost regressor using GridSearchCV to find the best model.
• Built a web app using Streamlit.

Run it

Locally

• Download the project and extract it
• Run Python 3.9 on a terminal
• Navigate into the folder you extracted
• Install the requirements : pip install -r requirements.txt
• Then : streamlit run src/app/main.py

Code and ressources used

Python version 3.9

Packages: pandas, numpy, sklearn, matplotlib, seaborn, beautiful soup, flask, streamlit, joblib

Libraries : Python library for speedrun.com API

Scraper Repository: https://github.com/kaparker/tutorials/blob/master/pythonscraper/websitescrapefasttrack.py

Scraper Article: https://blog.lesjeudis.com/web-scraping-avec-python (french)

Flask Productionization: https://www.analyticsvidhya.com/blog/2020/04/how-to-deploy-machine-learning-model-flask/

Data collection

Web scrapping

Tweaked the web scrapper above to scrape 2000+ games from tasvideos.org. With each game the following :

• Time of the best TAS
• Emulator the TAS has been made with

Speedrun.com API

Used the python library to use speedrun.com API (above) and extract 3800+ runs with the following :

• game name
• category
• platform
• engine the game has been made with
• developpers
• publishers
• released year
• number of runners for each game

Data cleaning

After scraping the data, I needed to clean it up so that it was usable for our model. I made the following changes and created the following variables:

• put the game and category in the same column
• calculate the age of game from released year
• removed runs with WR better than collected TAS time
• created a variable to control outliers : time_difference
  • removed outliers for runs with time_difference > 20
• added columns coding age and number of runners for each game to categorical variables

EDA

It revealed that number of runners were a key feature to predict the best possible time.

Model building

First, I transformed the categorical variables into dummy variables. I also split the data into train and tests sets with a test size of 30% (studies shows the best repartition is somewhere near that).

I tried different models :

• Simple linear regression
• Multiple linear regression
• Multiple linear regression with significative feautres (p-values < 0.2 on linear regressions)
• Lasso Regression
• Ridge Regression
• Random Forest Regressor
• Gradient Boost Regressor

Model performance

The best model was the Gradient Boost Regressor with a MAE of ~388 seconds on test set.

Productionization

In this step, I built a flask API web app prototype using the relevant link in reference on a local webserver. The prototype allows the user to type the name of the game, choose the category and returns the link of the world record, its time and the estimated best time.

I decided to do the actual app using Streamlit.

The online version is coming soon.