From-Scratch Structure-Aware Learning for Autonomous DJ Transition Planning

This proposal investigates whether a dedicated music-domain model trained from random initialization can support autonomous DJ performance through shared structural representations, directional section compatibility, cue ranking and interpretable gain/EQ control. The long-term research question includes temporal listener engagement, track selection, sequencing and session continuity; the initial experiment isolates structure-aware transition planning and source-preserving rendering. The proposed compact encoder will be trained on legally usable music and commissioned demonstrations, with deterministic DSP executing bounded controls. Evaluation compares handcrafted, randomly supervised, in-project self-supervised and separate-model baselines using leakage-resistant splits, objective timing/artifact measures and blinded listening tests. A conditional behavioral extension models timestamped exit, skip and replay observations, distinguishing monotonic survival from nonmonotonic section-engagement profiles. Engagement can rerank only musically feasible actions and is not treated as musical quality. The work proposes an auditable training lineage and experimental pathway rather than claiming universal autonomy or completed results. Feasibility depends on cleared, aligned data, independent evaluation and measured resource limits.