Papers

Symbolic regression seeks to discover closed-form mathematical expressions that accurately describe observed data, yet standard genetic-programming approaches frequently suffer from premature convergence and diversity collapse as populations stagnate on local optima. AHISTER introduces a dynamic multi-branch evolutionary framework — a mathematical colosseum in which independent evolutionary branches first develop in isolation, then enter the arena under tournament-based asymmetric partitioning into intensification and diversification paths. A coordinating branch aggregates a real-time Hall of Fame from asynchronous sub-branches; a dual stagnation clock governs termination at both local and global levels.

↓ Read current draft (PDF)

Distributed evolutionary search across heterogeneous compute clusters poses two coupled challenges: where should newly generated work execute, and how much effort should each unit of work consume? DEBOHAC is a holonic cluster scheme designed for symbolic-regression workloads in which independent branches and their recursive sub-branches must be distributed across an arbitrary-depth tree of cooperating instances. Capacity flows upward as periodic situation reports that aggregate hierarchically; budget flows downward as the authority to evolve. Each instance independently runs a three-step orchestration cycle that instantiates the principle of subsidiarity at every level.

Full-fidelity evaluation of evolved candidates is the dominant cost in symbolic regression at scale. CRESCENT introduces approximate-evaluation gating: cheap binned evaluations decide whether a candidate deserves a full-fidelity score, redirecting compute toward the most promising regions of the search space. We describe the gating mechanism, the binning prerequisite, and the experimental framework for isolating its contribution against an SRBench-grounded baseline.