# Porting plan (from SymbolicRegression.jl concepts)


This document maps major features/design choices to Rust equivalents and notes feasibility.


## Search & representation
- **Representation:** tree/graph AST with typed ops. Rust enums for ops and nodes. Feasible (done in MVP).
- **Search:** GP with tournament selection + crossover/mutation, plus Pareto archiving. Extensible to age-fitness, lexicase.
- **Constants:** LM/variable projection via `argmin`/`nalgebra`. Start with LM; later add autodiff for local gradients.


## Evaluation performance
- **Julia’s fused loops & SIMD:** replicate with `std::simd` and hand-rolled evaluators. Bench and specialize hot ops.
- **Parallelism:** use `rayon` for population eval; thread support optional in WASM (SharedArrayBuffer route).


## Simplification & pruning
- **Local rules:** identities (e.g., `x+0→x`, `x*1→x`, `sin(0)→0`).
- **Global rewriting:** integrate `egg` as optional feature for equality saturation + cost-based extraction.


## Multi-objective
- Maintain Pareto front on (error, complexity). Provide knobs for complexity penalty and max size/depth.


## Export & interop
- **Python:** PyO3 module `symreg_rs` with sklearn-like API (`fit`, `predict`, `score`).
- **WASM:** thin bindgen façade for browser demos; avoid heavy deps.
- **SymPy/LaTeX:** stringifier to SymPy code, plus LaTeX pretty-printing.


## Test & benchmark
- Reuse standard SRBench datasets where licensing permits; include Friedman1 synthetic.
- Add criterion benches for evaluator and variation operators.


## Milestones
1. MVP (this scaffold): scalar eval, GP loop, Python & WASM hello world.
2. Vectorized eval + `rayon` parallel pop eval; constant fitting.
3. Simplifier + export; sklearn-compatible estimator.
4. E-graph integration; advanced search strategies; docs + notebooks.