Life-long learning in game development processes

In the game industry, software projects extend over several years: for instance, a typical AAA game is developed for 3 to 5 years. To make the development process easier for the developers, tools are put at their disposal to help with, for instance, the artistic creation process or code integration. Those tools are based on machine learning (ML) models and learning processes, trained on historical data.
While such ML models might work accurately right after training, their predictions might lose relevance as time goes by, since the project's development objectives change over time, becoming progressively different from those at the time of model training. The phenomena of models and dataset aging are referred to as concept drift in the literature.
Through this research, we want to study concept drift in the game industry using the years of historical data and tool performance Ubisoft has gathered over time. Better understanding of concept drift would allow game tool creators to decide strategically when to update learning processes, and with which dataset, in order to maximize the performance (accuracy, precision, recall) of the tools over a long period of time.

Doriane Olewicki
Superviseur universitaire: 
Sarath Chandar Anbil Parthipan;Bram Adams
Partner University: