Content-based Game Classification with Deep Convolutional Neural Network

An attempt to better understand games from their visual aesthetics and game mechanics. I'm using gameplay videos to extract images that I can use with CNN to extract generic visual descriptors.  With these features, I can train classifiers to recognize different categories of games such as RTS, FPS, platformers, etc. The project is still in early phases put you can play around with the early demo here

Ropossum: A Mixed-Initiative tool for Evolving Playable Content for Physics-Based Games

In collaboration with Mohammad Shaker. This project investigates the use of different machine learning and procedural content generation techniques to generate endless variations of playable content for a clone of the popular physics-based puzzle game Cut the Rope. The project also features a mix-initiative design tool in which a partially designed level by a human (designer or player) can be automatically completed and checked for playability. The tool also provide real-time visual feedback about designs. Demos and showcases can be found here.

Adaptive Game Content Generation

The goal of this project is to automatically generate content that is tailored to specific playing style. The project demonstrates several cutting-edge types of AI in order to capture, recognize and model player experience. It also explore the use of procedural content generation techniques to automatically adjust game content generation and optimize chosen player experience. Computational models of players experience that map game, gameplay and behavioral features to players affects are constructed from crowd-sourced data and used as evaluation functions for game content generation. This project is supervised by Georgios Yannakakis, and Julian Togelius. Demos and showcases can be found here.

Designed by Mohammad Shaker, Strong Emotions ©, 2015