Adaptive Music Generation for Computer Games

This dissertation explores a novel approach to gamemusic that addresses the limitations of conventional gamemusic systems in supporting a dynamically changing narrative. In the proposed approach, the music is generated automatically based on a set of variable input parameters corresponding to emotional musical features. These are then tied to narrative parameters in the game, so that the features and emotions of the music are perceived to continuously adapt to the game’s changing narrative. To investigate this approach, an algorithmic music generator was developed which outputs a stream of chords based on several input parameters. The parameters control different aspects of themusic, including the transitionmatrix of aMarkovmodel used to stochastically generate the chords, and canbe adjusted continuously in real time. A tense first-person game was then configured to control the generator’s input parameters to reflect the changing tension of its narrative—for example, as the narrative tension of the game increases, the generatedmusic becomesmore dissonant and the tempo increases. The approach was empirically evaluated primarily by having participants play the game under a variety of conditions, comparing them along several subjective dimensions. The participants’ skin conductance was also recorded. The results indicate that the conditionwith the dynamically variedmusic described above was both rated and felt as the most tense and exciting, and, for participants who said they enjoy horror games and films, also rated as the most preferable and fun. Another study with music experts then demonstrated that the proposed approach produced smoother musical transitions than crossfades, the approach conventionally used in computer games. Overall, the findings suggest that dynamic music can have a significant positive impact on game experiences, and that generating it algorithmically based on emotional musical features is a viable and effective approach.