What benchmarks can teach us about AI-assisted game design

Speaker Name: 
Vanessa Volz
Speaker Title: 
Postdoctoral Research Associate
Speaker Organization: 
Queen Mary University
Start Time: 
Monday, March 4, 2019 - 11:15am
End Time: 
Monday, March 4, 2019 - 12:15pm
Location: 
E2-506
Organizer: 
Jim Whitehead

 

 

Abstract: AI-assisted game design approaches, such as procedural content generation (PCG), are popular in research and have also garnered increasing interest in the game industry. However, many approaches are not fully understood, making it difficult to assess their success and potential application in other games. In this seminar talk, I will present recent efforts to alleviate these problems which are based on the game-benchmark for evolutionary algorithms (GBEA).

The GBEA is built from a taxonomy and survey of automatic evaluation approaches of games and game content. I will thus first present the taxonomy, before going into more detail on the four single- and multi-objective function suites that are part of the GBEA. The function suites contain diverse sets of problems resembling those proposed in previous research on AI-assisted game design. The main part of the talk will then focus on the insights obtained through the benchmark and what recommendations result from it. I will further discuss how these results can be generalised as well as present other directions for future work in AI-assisted game design.

Bio: Vanessa Volz is a post-doctoral research associate at Queen Mary University London, UK, with focus in computational intelligence in games. She received her PhD in 2019 from TU Dortmund University, Germany, for her work on surrogate-assisted evolutionary algorithms applied to game optimisation. She holds B.Sc. degrees in Information Systems and in Computer Science from WWU Münster, Germany. She received an M.Sc. with distinction in Advanced Computing: Machine Learning, Data Mining and High Performance Computing from University of Bristol, UK, in 2014.

Her main research interests are two-fold: One is the analysis of game optimisation problems (such as search-based procedural content generation or automatic tuning / balancing) in terms of their complexity, uncertainty and other patterns. The other interest is the development of algorithms specifically suited for these problems, such as surrogate-assisted evolutionary algorithms that are able to handle non-symmetric noise and uncertainties.