Skip to content

Latest commit

 

History

History
41 lines (21 loc) · 2.79 KB

readme.md

File metadata and controls

41 lines (21 loc) · 2.79 KB

This is a copy of a private repository, and so does not contain the project commit history.

Prediction of the Turtle Strategy in DOTA 2

This repository contains the source code written for completion of my undergraduate dissertation, to recognise occurences of the turtle strategy in DOTA 2. The paper is available to read on my personal website.

The code provided has been adapted for use with distributed GPU computing. Some minor configuration may be required for use with CPU instead.

Inspiration

The model was built after the architecture presented in Adam Katona's brilliant paper on death prediction, which is well worth a read.

Architecture

The strategy is modelled with a deep, feedforward neural network, constructed with tensorflow 2.3.0:

Rough architecture sketch

Each hero is represented by 21 input features. Each hero's features is fed to a single subnetwork. Since the representations of heroes should be calculated in the same way, the weights of these hero subnetworks are shared - and so the hero representation subnetwork is called the "shared" subnet.

Each hero's 21 outputs from the shared subnetwork are concatenated into one long vector, which is fed into a final subnetwork with 10 outputs. Each output is regression between 0 and 1, predicting the probability that each hero is currently attempting to employ the turtle strategy.

Results

The network was trained with a small and highly imbalanced dataset of 37 professional Dota 2 matches, and evaluated at an average precision of 9.04%, and a maximum f1 score of 13.13%.

Data

The model reads either one or a set of CSV files, where columns are features, and rows are game state readings for a single game tick.

This data was extracted from DOTA 2 replay files using the Clarity parser, my extension of which is provided in this repository. What is not provided here is the labelling process, which is manual. CSVs must end with 10 label columns, isTurtling0, isTurtling, ... isTurtling9.

The pre-parsed dataset used during my dissertation project is available here.

Special Thanks

Deepest thanks go to my supervisors Prof Anders Drachen and Prof James Walker, for their continuous, invaluable support and enthusiasm throughout.

A huge thank you also to for the continuous help and expert insights offered by Dr Athanasios Kokkinakis, Dr Simon Demediuk, and especially Alan Chitayat, who helped immensely without hesitation.

Lastly, special thanks go to my wonderful parents and girlfriend, for their support during this project and the during difficult times surrounding it.