LIP6 1997/42
- Reports
Construction d'un Agent Rationnel jouant du Jazz - G. Ramalho
- 306 pages - 01/13/1997- document en - http://www.lip6.fr/lip6/reports/1997/lip6.1997.042.pdf - 1,752 Ko
- Contact : Geber.Ramalho (at) nulllip6.fr
- Ancien Thème : APA
- Keywords : automatic music accompaniment, intelligent agents, case-based reasoning, computer music
- Publisher : Jean-Luc.Mounier (at) nulllip6.fr
Artificial intelligence programs have reached a certain success in performing some classes of tasks such as classification and prediction. However, the design and creation tasks, particularly the artistic ones, seem to be more problematic. In this context, our work has been devoted to the study of the strengths and limitations of current AI techniques in modelling the knowledge necessary to build a machine capable of creating and playing jazz bass lines in real time. Based on the notion of rational agent, we have proposed a original problem solving model that integrates different kinds of knowledge and reasoning, in particular case and rule based ones. According to our model, the task of a jazz player agent can be divided into three main steps, which occur continuously. First, by means of a particular harmonic analysis, the agent establishes the "temporal segment" within which a melodic phrase must be created and played. Second, the agent activates an ensemble of “Potential ACTions” (PACTs), such as "play with high dissonance" and “play a lot of notes”. These PACTs can be seen as musical instructions concerning different musical facets (e.g. dissonance and density) of the melodic phrase to be created and played. Third, the agent uses its "Musical Memory" (a case base containing melodic previously played by human musicians), to retrieve the fragment that fits the best the activated PACTs. To validate our model, we have implemented the system ImPact, containing 267 cases and more than 80 production rules. The musical results have been as good as those of the two “best” current programs which can generate bass lines (Band-in-a-box and NeurSwing). The main advantage of our model is that additional knowledge can be easily incorporated to improve the system’s musical performance.