Case Study: Die Roller Demo — Emergent Entropy in Structured Systems
Foreword \\ This was originally built as a precursor to an RL randomness study comparing true‑entropy agents (True Bots) and classical RNG policies. Hosting limits paused that track; this demo summarizes the core observation used to inform it. See True Bots.
Summary \\ Over 100k+ d20 rolls (5 dice/call), we compare PRNG vs ERIS (quantum‑native entropy). Tracking χ², Uniformity Delta, variance, and plateaus reveals ERIS preserves variance and resists local minima; PRNG stabilizes into uniform mimicry. ERIS explores edges; PRNG hugs the center.
Key Terms
- Structured Randomness \\ high unpredictability with coherent long‑range patterning (e.g., fractal bitstreams)
- Entropy‑Structure Dichotomy \\ high measured entropy coexisting with visible structure
- Local Minima Resistance \\ avoidance of long uniform plateaus (stagnation)
- Variance Preservation \\ sustained χ²/Uniformity Delta to maintain exploration
- Pre‑processing Sensitivity \\ whitening can mask structure yet underlying dynamics persist
Method
- 5× d20 per API call; matched totalRolls across PRNG and ERIS
- Metrics: χ² (dof=19), Uniformity Delta, histograms, mean/median/stdev
- Intervals: analyze trends over 3k+ windows; observe divergence on ~5k–7k+
- Note \\ reruns used identical configs; only entropy source differs
Results and Interpretation
- Early phase (≤5k) \\ both oscillate similarly; divergence begins ~5k–7k
- Minima resistance \\ visible ~7k; stabilizes ~13k+; confirmed ~14.5k+
- ERIS χ² range \\ repeatedly breaches upper bounds (e.g., 114 at 103k; >160 at 148k); PRNG often stays 15–30
- Uniformity Delta \\ ERIS ~2× PRNG scale; sustained climbs and cycles
- Harmonization \\ by 103k, ERIS mean/median ≈10, stdev ≈5; exploration without collapse
Why this matters for RL and Game Systems
- Exploration strategy \\ ERIS prevents early convergence; sustains novel state traversal
- Reward shaping \\ adds resonance, not flat jitter; smoother transitions across cliffs
- Policy mutation \\ multiscale variability avoids dead‑end loops
- Procedural gen \\ non‑stale worlds and narratives (resists repetitive attractors)
Related case study: Entropy Microdemo; see also Meta Llama 3.1 post‑training and Rejection Sampling discussion [link](https://medium.com/@jkabrit/metas-post-training-pipeline-for-llama-3-1-e6777801c0a1).
Resources
Back to Game Dev Corner