Quick Start Guide

CLICK ON ME FOR VIDEO GUIDE!

1. Install Python 3 (preferably latest version)
   • Install wheel package - pip install wheel if not installed
2. Clone repo
3. run pip install -r requirements.txt in repo directory to install required packages
4. copy gamestate_integration_CSGOPredictor.cfg file to Steam\steamapps\common\Counter-Strike Global Offensive\csgo\cfg
5. run python MainApp.py plus any arguments of your choice in command terminal or equivalent from location of this repo

Optional - run gui.py in another terminal while MainApp.py is running to display dynamic prediction bar

The program will begin making predictions once you begin spectating a match in CS:GO.

Requirements

• Python 3 (preferably latest version)
• wheel python package
• All packages in requirements.txt (install wheel before requirements.txt)

(And Counter-Strike: Global Offensive, of course!)

Features

★ Prediction

The program print the live round prediction in the terminal. Prediction format is [CT Win%, T Win%]

• The program also writes each prediction to predictions.txt in repo directory.

★ Pause-and-Play

Hold Esc key while terminal window is active to pause the program! (currently only works when run using Anaconda)

★ GUI

Run gui.py in another terminal while MainApp.py is running to display dynamic prediction bar

★ Command Line Arguments

• -w = disable Welcome Message
• -p = disable Pause-and-Play functionality
• delay X = delay predictions by X seconds

CSGOPredictor - An Overview

CSGOPredictor is a python program that generates live round winner predictions of CS:GO Competitive matches.

How it works

A simple 3 step process

1. When a match is live, snapshots of the the round in play, containing large amounts of precise data on round & players' status, are generated & captured using the gsi_pinger module through CS:GO's in-built Game State Integration functionality.

2. Each snapshot is cleaned and parsed using the snapshot_parser module, resulting in the creation of an array of 23 attributes to be used by the predictive model to generate probability predictions. Attributes include:

• Round Data - Map, Time Left, Bomb Plant Status
• Player Data - T/CT Players Alive, T/CT Total Health, Weapons, Utility

3. Finally, MainApp.py runs the pre-trained Logistic Regression model to generate probability prediction for round at that particular point in the round.

• Prediction is in the form of an ordered duo of probabilities - first for CT win % and second for T win %.
  • Example: [79.21, 20.79], indicating a 79.2% win probability for CTs & 20.8% win probability for Ts
• The prediction is printed in the terminal as well as written to a text file in the parent directory (for use by other applications, such as gui.py which displays the predictions as a dynamic bar chart)

Metrics

The Predictive Model used in this program is

• a Logistic Regression model
• trained on this dataset, which contains 122,411 snapshots of from high level tournament play in 2019 and 2020.
• The dataset is a pre-processed version of the dataset released by SkyBox.gg as a part of their AI hackathon.
• The model was trained on all 122k+ snapshots, with 90+ attributes used out of the 97 present in the dataset. Some of the attributes were combined to create 23 final attributes which the model uses to make predictions.

Calibration Plot:

A calibration plot is a line-and-scatter plot which compares the observed probabilities of an event versus the predicted probabilities. A well calibrated predictor is one where results that are predicted with an X% probability do indeed occur X% of the time.

As the primary function of this program is to generate accurate probabilities, callibration is the key metric for success (not accuracy)

Accuracy:

Acknowledgements:

1. Christian Lillelund and Skybox.gg for data used to train the predictive model
2. mdhedelund for their CSGO-GSI Repo, which was used in this repo's gsi_pinger module to interface with CS:GO's GSI
3. mlrequest for their sklearn-json package, used to import the predictive model in MainApp.py

Special Thanks to My Research Guide Dr. Deepak Joy Mampilly, and my Business Analytics professors Dr. Kumar Chandar S & Dr Manu K S, for their guidance & support.