# AI MusiMuse # 013_COMPOSITION_SEARCH.md Version 1.0 --- # Purpose The Composer should never generate the first possible continuation. Instead it should explore multiple possible futures and choose the one that best matches the author's musical style. Composition therefore becomes a search problem. --- # Philosophy Traditional music generators operate as Current Context ↓ Next Token ↓ Next Token ↓ Next Token AI MusiMuse operates as Current State ↓ Possible Decisions ↓ Future Simulation ↓ Evaluation ↓ Best Decision Generation is therefore planning rather than prediction. --- # Search Loop ``` Current State ↓ Generate Candidates ↓ Simulate Futures ↓ Evaluate ↓ Select Best ↓ Update State ↓ Repeat ``` --- # Current State At every generation step the Composer knows Current section Current energy Current density Current harmony Current texture Current motifs Elapsed time Remaining duration User constraints Style probabilities --- # Candidate Generation The Composer generates multiple possible decisions. Example ``` Add Bass Increase Energy Return Theme Create Break Introduce Variation Change Harmony Do Nothing ``` Typically ``` 5–20 candidates ``` per step. --- # Future Simulation Each candidate is expanded into a hypothetical future. Example ``` Decision ↓ New State ↓ Next Decisions ↓ Future Composition ``` The simulation depth is configurable. --- # Evaluation Every simulated future receives a score. Example ``` Style Score 0.91 Coherence 0.84 Novelty 0.73 Constraint Satisfaction 1.00 Overall 0.89 ``` --- # Evaluation Components Overall score consists of weighted terms. ``` Style Match Structural Consistency Energy Continuity Density Continuity Harmony Consistency Motif Development Novelty Constraint Satisfaction ``` Weights are configurable. --- # Style Score Questions ``` Would the author probably do this Has this transition occurred before Does it resemble learned habits Does it violate the Style Model ``` --- # Novelty Score Novelty prevents copying. Examples ``` Exact repetition ↓ Penalty Small variation ↓ Preferred Completely random ↓ Penalty ``` The highest score belongs to balanced originality. --- # Constraint Score Examples ``` Ambient only No percussion Long intro Minor mode Low energy ``` Constraint violations reduce the score. --- # Beam Search The default search algorithm is Beam Search. Example ``` Width = 8 Depth = 5 ``` Only the strongest candidates survive each expansion. --- # Alternative Algorithms The architecture supports Beam Search Monte Carlo Tree Search A Best First Search Sampling Hybrid Search The Composer is independent of the search strategy. --- # Search Tree ``` State ├── Decision A │ ├── Future A1 │ └── Future A2 ├── Decision B │ ├── Future B1 │ └── Future B2 └── Decision C ``` Only the best path is selected. --- # Decision Memory During one generation the Composer stores Visited states Repeated motifs Previous decisions Current plan This prevents loops. --- # Replanning The Composer may reconsider. Example ``` Initial Plan ↓ Unexpected evaluation ↓ Choose different branch ``` Generation is adaptive. --- # Randomness Random Seed influences candidate ordering tie breaking creative variation The same seed always produces identical output. --- # Determinism Given Style Model Seed Constraints the Composer always produces the same composition. --- # Future Neural Integration In future versions Candidate Generation may be neural. Evaluation may be neural. Search remains explicit. This keeps generation explainable. --- # Explainability Every generated decision can be explained. Example ``` Decision Return Theme Reason Occurs in 82% of similar compositions Energy curve predicts thematic return Matches user constraints ``` Generation should never be a black box. --- # Design Principles Planning before generation. Multiple futures explored. Search independent of AI model. Evaluation independent of renderer. Deterministic generation. Explainable decisions. --- # Long-Term Goal The Composer should behave like a human composer, who mentally explores several possible continuations, rejects weaker ideas, and develops the strongest one into a finished composition.