Skip to content

QUICK_START_PHASE5.md

Latest commit

 

History

History
117 lines (84 loc) · 3.68 KB

QUICK_START_PHASE5.md

File metadata and controls

117 lines (84 loc) · 3.68 KB

Phase 5 Fixes - Quick Start Guide

What Was Fixed

Your Phase 5 training revealed that agents couldn't learn because the world (300×300) was too large for them to discover resources during their exploration phase. I've implemented three critical fixes:

  1. World size reduced: 300×300 → 150×150
  2. Resource counts adjusted: Scaled to maintain density
  3. Spawn location optimized: Agents now spawn near resource clusters

Quick Test (5 minutes)

cd soliter-develop
python scripts/train_soliter.py --cycles 50 --output-dir experiments/phase5_quick_test

What to look for in the output:

  • First consumption should appear before cycle 20
  • "Consumed" column should show numbers (not just dots)
  • By cycle 50, you should see 10+ consumption events

Expected Results

Before (Your Training Log)

  • 400,000 ticks, 200 cycles
  • 7 total consumptions (0.00175% rate)
  • Agent explored only 23% of world
  • No learning occurred

After (With Fixes)

  • 100,000 ticks, 50 cycles
  • 50-150 consumptions (>10% rate)
  • Agent explores >50% of world
  • Actual learning visible

Files Modified

  1. configs/default.yaml - World size and resource counts
  2. scripts/train_soliter.py - Spawn and respawn logic

No other changes needed - all your Phase 4 work (memory consolidation, EWC, drive system) is intact.

Validation Checklist

After running the 50-cycle test, check:

  • Console shows consumption events (food/water/heat)
  • Satisfaction values increase over time
  • Agent position varies widely (not stuck in one spot)
  • Fewer deaths from the same cause repeatedly
  • Buffer size grows steadily (experiences being collected)

Next Steps

If Test Passes ✅

Run longer experiment:

python scripts/train_soliter.py --cycles 200 --output-dir experiments/phase5_validation

Then analyze with:

python scripts/plot_training.py experiments/phase5_validation/training_*.json

If Test Fails ⚠️

Low consumption (<5 events):

  • Try even smaller world (100×100)
  • Increase initial action_std to 1.0

Resource discovery but no consumption:

  • Check drive_system activation in logs
  • Verify gradient sensor responses

Consumption but no learning trend:

  • Check if rewards are being collected
  • Verify PPO updates are occurring

Understanding the Fix

The problem wasn't your code - it was a scale mismatch:

Old Configuration:
  World: 300×300 (90,000 units²)
  Agent exploration: ~60 unit radius (11,300 units²)
  Coverage: 12.5%
  Result: Most resources unreachable → no learning signal

New Configuration:
  World: 150×150 (22,500 units²)  
  Agent exploration: ~60 unit radius (11,300 units²)
  Coverage: 50%
  Result: Most resources discoverable → learning possible

Why This Matters for Consciousness Research

Your project aims to demonstrate consciousness prerequisites through continual learning. But you can't test consciousness emergence without functional learning first. These fixes ensure:

  1. Agents can learn - Resources are discoverable
  2. Drive system functions - Internal rewards get satisfied
  3. Memory consolidates - There's something to remember
  4. Context integrates - Experiences across cycles connect

Without these fixes, you were testing memory systems in a "sensory deprivation" environment. Now the agent can actually engage with its world.

Contact

If you encounter issues, the key diagnostic is consumption rate:

  • <1% = Still broken (try smaller world or higher action_std)
  • 1-10% = Marginal (might work with more cycles)

  • 10% = Working! (proceed to long-horizon testing)

Good luck! The mechanical systems are solid - they just needed a world they could actually explore.