VIVIA — MeMLP Neural Network

Overview

MeMLP (Modular embedded Multi-layer Perceptron Model) is the neural network stack embedded directly in the VIVIA engine. It runs 100% on CPU, in-process, with no GPU, no cloud, no external runtime.

Architecture

Module	Shape	Task
vegetation	8 → 24 → 16 → 4	flower / fern / stick / pebble placement
biome	8 → 12 → 9	biome classification (9 world biomes)
texture	8 → 12 → 6	procedural texture-style selection
motion	8 → 16 → 6	creature animation parameters (sigmoid output)

Total: ~3,400 parameters · Checkpoint size: ~14 KB JSON · Inference: 0.01 ms

Feature Vector (8 dimensions)

[0] terrain_height     — normalised 0..1
[1] terrain_slope      — 0 flat, 1 vertical
[2] biome_temperature  — from embedded NASA climatology
[3] biome_humidity     — from embedded NASA climatology
[4] water_distance     — proximity to water
[5] vegetation_count   — nearby plant density
[6] light_level        — sun exposure
[7] noise_seed         — procedural variation

Training Pipeline

┌─────────────────┐
│  Synthetic Data │──┐
│  (heuristic)    │  │
└─────────────────┘  ├──► MeMLP Training Loop ──► Checkpoint Save
┌─────────────────┐  │         (online, background thread)
│ Open-Meteo API  │──┘
│ (real weather)  │
└─────────────────┘

• Online training: background thread, 100 samples/epoch
• Loss function: cross-entropy (vegetation, biome, texture) + MSE (motion)
• Optimiser: SGD with gradient clipping, weight decay, NaN sanitisation
• Learning rate: 0.01 with decay every 1000 epochs
• Legacy migration: old single-hidden-layer checkpoints auto-upgraded

Motion Head (Creature Animation)

The motion module outputs 6 continuous parameters (0..1, sigmoid) that drive live creature animation:

Channel	Parameter	Effect
0	gait_amp	Leg swing amplitude
1	gait_freq	Step frequency multiplier
2	breath	Torso breathing depth
3	tail_wag	Tail sway amount
4	head_dip	Grazing/head-lowering
5	hop	Vertical bounce (birds, fleeing)

Player Preference Learning

The AI tracks player block placements and adjusts training targets to match the player's aesthetic preferences. This means the vegetation patterns adapt to each player's style over time.

Online Training (Open-Meteo)

Real-world weather data from 8 global locations (Warsaw, Nairobi, Dubai, Amazonas, Reykjavik, Beijing, Sydney, London) feeds into training as climate features. Falls back to synthetic data when offline.

Tests

12 MeMLP-specific tests covering:

• Forward pass validity (probability distribution, sigmoid range)
• Deterministic initialisation
• Training loss decrease
• Legacy checkpoint migration
• Motion head independence (sigmoid vs softmax)
• NaN/Inf survival
• Roundtrip serialization