A pair of state transition models was constructed based on physiological and self-report data:
• The first model tries to detect the listener’s current affective state.
• The second model chooses the direction most likely to induce the goal state.

Conclusions

• changes in music parameters can be correlated to affective response to music
• Markov chains are a useful tool for constructing a predictive listening system
• specific observations about mapping layering/complexity to arousal/valence:
• engaged listeners tend to stay engaged
• annoyed listeners tend to stay annoyed
• soothed listeners tend to stay soothed, but also easily bored or engaged
• bored listeners tend to become interested by any change in parameters
• annoyed listeners are more likely to be engaged if first induced to boredom

Future Work
• improve accuracy of predictions by incorporating more user data
• improve affective state predictions using additional affective signals
• apply affect-parameter mapping to algorithmic composition
• use machine learning to customize predictions to individual subject