# MLE Competitive Model

A high-dimensional sparse distributed memory system with energy-based dynamics and knowledge-based question-answering capabilities.

## Architecture

- **Vector size**: 4096 bits (sparse, ~5% active = 205 bits)
- **Memory**: Sparse Address Table with deterministic hash encoding
- **Knowledge base**: 1848 facts, 1106 words, 53 sequences
- **Encoding**: MD5-based deterministic sparse vectors
- **Inference**: Symbolic reasoning with Hamming similarity

## Benchmarks - Ultimate Model (v3)

| Task | Accuracy | Details |
|------|----------|---------|
| **Question Answering** | **62.8%** | 49/78 correct |
| **Analogy Solving** | **67.9%** | 19/28 correct |
| **Sequence Completion** | **93.3%** | 14/15 correct |
| **Word Retrieval (noise 15%)** | **100%** | 100/100 correct |
| **Generalization** | **100%** | 20/20 correct |
| **Robustness (noise 50%)** | **100%** | Up to 50% bit corruption |
| **Overall** | **84.8%** | Weighted average |

## Corpus Coverage

- **Animals**: 185 animals across 5 categories
- **Plants**: 175 plants across 6 categories
- **Locations**: 150+ cities across 15 countries and 5 continents
- **Vehicles**: 10 vehicles with parts and actions
- **Professions**: 20 professions with workplace associations
- **Food**: 55 food and drink items
- **Body**: 36 body parts with functions
- **Colors**: 29 colors with associations
- **Emotions**: 20 emotions with opposites
- **Time**: Full temporal sequences (days, months, seasons, ordinals)
- **Tools**: 20 tools with functions
- **Music**: 17 instruments and 12 genres
- **Sports**: 22 sports
- **Weather**: 14 weather types
- **Clothing**: 20 clothing items
- **Technology**: 19 devices
- **Celestial**: 20 celestial bodies

## Files

- `mle_ultimate_model.npz` - Word vectors (1106 × 4096 binary)
- `mle_ultimate_facts.json` - Complete knowledge base (1848 facts)
- `mle_ultimate_results.json` - Benchmark results
- `mle_ultimate_config.json` - Model configuration

## Usage

```python
import numpy as np

# Load model
data = np.load("mle_ultimate_model.npz", allow_pickle=True)
words = data['words']
vectors = data['vectors']

# Each vector is a 4096-bit sparse binary vector
# ~205 bits active (5% sparsity)
```

## Supported Query Patterns

- "what is X" → category (is_a)
- "what has X" → possessor (has)
- "what can X" → agent (can)
- "where is X" → location (in/lives_in)
- "what color is X" → color (is/has_color)
- "what is opposite of X" → antonym (opposite)
- "what is before X" → predecessor (before)
- "what is X made of" → material (made_of)
- "what is X used for" → function (used_for)

## Model Versions

| Version | Facts | Words | Overall |
|---------|-------|-------|---------|
| v1 (mle_best) | 283 | 334 | **97.2%** |
| v2 (mle_advanced) | 933 | 705 | **68.2%** |
| v3 (mle_ultimate) | 1848 | 1106 | **84.8%** |

v3 offers the largest knowledge coverage with strong robustness to noise.