Description

The following project was completed for my Individual Research Project in my final semester at the University of Gloucestershire. This module was essentially my undergraduate dissertation. This project won the award for the best research project in 2023. My goal was to create a Convolutional Neural Network (CNN) that would be able to accurately predict the percentage of positive reviews a game would receive on Steam with the only input being a single static gameplay image. Such a tool would be extremely valuable to any game developer as they would be able to predict player feedback during development. In theory thousands of images from the gameplay loop could be inputted into the network and the results plotted to show strong and weak areas of the game.

All code for this project was written in Python. To collect my training data I wrote a webscraper to label and download approximately 400,000 images from the Steam store. I then used various libraries such as Tensorflow to build my network architecture and train my model. The most successful model used transfer learning of the Xception model weights which was trained on ImageNet images. The Xception weights were then fine tuned to better fit the input data and model expectations.