Active Inference in Action Using an Embodied Agent for Co Creative Musical Interaction

Interactive music systems for ensemble co-creation commonly rely on heuristic architectures that lack both a unified theoretical foundation and interpretability. We address this limitation by formulating the system design within the Active Inference framework. We introduce the Active Embodied Musical Intelligence Agent (AEMIA), a computational model that frames musical interaction as a Partially Observable Markov Decision Process. In this formulation, perception corresponds to variational state estimation, while action selection emerges from minimizing Expected Free Energy (EFE), thereby balancing goal-directed behavior (pragmatic value) against information-seeking exploration (epistemic value). The agent’s aesthetic intentions are explicitly encoded as a prior over preferred sensory outcomes (\(\:{P}_{pref}\)), with a Mixture Density Network modeling the joint likelihood of multi-modal sensory inputs. In comparative experiments with musicians, AEMIA significantly outperformed its heuristic predecessor and other baselines, achieving a co-creativity score of 8.9/10. An ablation study confirmed the necessity of both the aesthetic prior and the epistemic drive for effective interaction. Qualitative feedback indicated that AEMIA’s actions were perceived as more intentional and conversational. This work establishes a rigorous, interpretable foundation for building co-creative musical agents.