Simulating Creativity in Jazz Performance

Geber L. Ramalho, Jean-Gabriel GanasciaAAAI 1994

Abstract

Jazz improvisation presents a unique challenge for artificial intelligence: the need to bridge the "gap" between a rigid chord grid and a soulful, spontaneous performance. This paper introduces a model that simulates this creative process by combining rule-based logic with Case-Based Reasoning (CBR). The architecture, centered on a "Reasoner" that manages "Potential ACTions" (PACTS), utilizes a "Musical Memory" of transcribed fragments to generate non-deterministic musical solutions. By incorporating internal states such as "Mood," the system moves beyond deterministic grammars to approximate the messy flexibility of human musicianship.

Introduction

In the early 1990s, AI researchers grappled with the paradox of creativity—a quality often defined by its resistance to formalization. In jazz, this is manifest in the distance between a basic lead sheet and the rich, expressive notes played by a musician. Previous AI approaches often fell into the traps of either pure randomness, which lacks structural integrity, or rigid grammars, which lack the "soul" and flexibility of live performance. This research proposes a middle ground: a system that models a musician’s behavior by integrating past experiences with current intentions within an evolving context.

System Architecture

The core of the proposed model is the "Reasoner" architecture, which operates in parallel with a "Context" module representing the external musical environment.

The Reasoner architecture and its interaction with the musical context.

The Reasoner is divided into three primary modules:

  1. Perception: Interprets external events and the performance of other musicians (e.g., the "Orchestra").
  2. Composing: The central engine that generates musical intentions and selects specific notes.
  3. Execution: Translates the composed intentions into actual performance data (e.g., MIDI output).

The PACTS Framework

To navigate the abstraction from intention to execution, the authors introduce PACTS (Potential ACTions). These are frame-like structures representing musical goals at varying levels of detail. A high-level PACT might be a vague intention like "play bluesy," while a "playable" PACT contains a fully specified sequence of notes. The improvisation process is viewed as a problem-solving task: refining a set of vague PACTS into a single playable action through three operators:

  • Delete: Resolves conflicting intentions based on the system's internal state.
  • Combine: Merges compatible goals (e.g., "swing" + "arpeggio").
  • Add: Fills in missing details by querying the Musical Memory.

Musical Memory and Case-Based Reasoning

Rather than relying solely on rules, the system utilizes Musical Memory—a repository of melodic and rhythmic fragments transcribed from real jazz recordings. Using Case-Based Reasoning (CBR), the "Composing" module searches this memory for "cases" that match the current chord, style, and mood. This allows the system to leverage human-level expertise and avoid the mechanical feel of purely algorithmic music.

Key Insights and Results

A significant feature of the model is the "Mood" Automaton. The system maintains internal states (e.g., "bluesy" or "in a hurry") that influence decision-making. For instance, if the system is "in a hurry" due to computational load, it may bypass complex planning and trigger a "default" action—a pre-learned "lick"—mimicking a human musician's reliance on muscle memory under stress.

The architecture supports non-determinism; given the same chord grid, the system can produce different valid musical outputs depending on its internal state and memory retrieval. By allowing retrieved memories to occasionally "short-circuit" rigid rules, the model captures the non-deterministic, expressive nature of jazz.

Conclusion

While the model provides a robust framework for simulating creativity, it faced limitations at the time of publication, particularly in the "Perception" module's ability to "hear" and react to its own output in real-time. Furthermore, the Musical Memory required intensive manual transcription and indexing. Future work points toward the automated acquisition of musical knowledge through machine learning and more sophisticated real-time replanning in the execution phase.

Summary updated

BibTeX
@inproceedings{ramalho1994simulating,
  author = {Geber L. Ramalho and Jean-Gabriel Ganascia},
  title = {{Simulating Creativity in Jazz Performance}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1994},
}
Figure