Lipson, et al
Eric Brochu, Hendrik Kuck, Nando de Freitas, falipson, ebrochu, kueck, nandog
Machine Learning for Game AI
Learning is potentially a powerful tool for computer game AI that is still largely unexploited. Games such as Creatures and the groundbreaking Black & White have made agent learning integral to games, but learning can also be bene_cial for more conventional games and can become a powerful tool for developers and researchers.
Machine Learning (ML) is the _eld of performing statistical analysis on data to learn the patterns therein. Recent research has explored the possibilities of applying ML techniques to agents in computer games. We are currently developing systems for training game agents from data generated by human players. By mining this data, we can use ML techniques to build models of human behavior. These in turn can be used as the basis of game agent actions: agents act by imitating the behavior of human players under similar circumstances.
Data sequences that capture di_erent styles of human play can be automatically discovered, grouped together and used to form models of di_erent play styles. The result is agents that play not in an optimal manner, but in a human-like style, making errors and applying strategies similar to humans. A happy side e_ect is that developers are not required to hard code rules for agent behavior. Coding AI becomes a matter of automatically building models based on data and extracting information from them. This can be done in a principled manner using well-understood ML tools. Applying the models to determine agent behavior is in principal as e_cient as the ad hoc methods 1 currently employed by AI developers. There are numerous implications of this technology: agents that learn to play like a speci_c human player; \autopilot" modes based on idiosyncratic behavior in persistent online games; opponents `trained' by human experts; and games that adapt to player behavior to keep play fun and interesting. Asher Lipson is an MSc student in Computer Science at UBC. He obtained his BScHons degree in Computer Science from the University of Witwatersrand, South Africa. Other than autonomous agents, he has interests in recon_gurable hardware, animation and DJing.
Eric Brochu is a PhD student in Computer Science at UBC. He received a BA in English Literature and a BSc in Computer Science from the University of Regina, Canada and is _nishing an MSc in Computer Science from UBC. His research include applying ML to trainable agents and multimedia visualization. Hendrik Kuck is an MSc student in Computer Science at UBC. He holds a diploma in Computer Science from University of Erlangen-Nurnberg, Germany. His research interests are in computer graphics, Machine Learning, and its applications to game AI. Nando de Freitas is an Assistant Professor of Computer Science at UBC. He holds BSc and MSc degrees in Electrical Engineering from the University of Witwatersrand, South Africa, and a PhD from Cambridge, England. His research interests include statistical multimedia, Monte Carlo methods, and Bayesian models of ML.
|
Sitemap 