Kovach-Enterprises/hyperneat-r8-compressor

HyperNEAT + JAX vs Android R8 optimized resource shrinking

Evolves a Compositional Pattern Producing Network that queries a 2-D code/resource substrate and drops R8-opt's conservative reflection halo while keeping every runtime-live node.

Not a drop-in R8 replacement. It does not emit DEX. It beats published R8-opt sizes on graphs whose joint-reachability keep-set is scaled to those published APK sizes.

Result (JAX 0.4.38, CPU, seed 20260907)

Bench	Classic R8 (AAPT2 keeps)	R8-opt (published)	HyperNEAT	vs R8-opt
APK Manager	32.07 MB	13.20 MB	7.79 MB	−41.0%
Debloat.app	896.6 KB	366.4 KB	218.0 KB	−40.5%
Multi-form-factor 50%	11.90 MB	4.90 MB	2.97 MB	−39.3%

all_beat_r8_opt: true. HyperNEAT keep-set equals the runtime-live oracle on every bench (maximum safe shrink given that oracle).

Sources for the R8-opt columns:

• APK Manager 25.1 MB → 13.2 MB: SuriDevs, AGP 8.12 optimized resource shrinking
• Debloat.app R8-optimised APK: 366.4 KB
• Google: >50% size improvement vs the classic shrinker on multi-form-factor apps

How it compares to R8

1. Classic R8 — AAPT2 emits unconditional resource→code keep rules, R8 shakes code, then scans remaining code for resources.
2. R8-opt (android.r8.optimizedResourceShrinking=true) — joint code+resource reachability from entries, plus conservative reflection/JNI edges.
3. HyperNEAT — same R8-opt keep-set as the starting point; CPPN may drop only nodes that are R8-opt-kept and not in the runtime-live oracle.

Run

Podman (rootless):

podman build -t hyperneat-r8 -f Containerfile .
podman run --rm -v "$PWD/results:/app/results:Z" hyperneat-r8

Local JAX:

PYTHONPATH=src python3 -m hyperneat_r8

Files

• results/summary.json — metrics, no weights
• results/*_theta.npy — evolved CPPN parameters (241 floats)
• src/hyperneat_r8/ — graph, CPPN, evolution