Implementation of the learning-functions brainstorm (L0–L5).

Architecture

[text]
nn learn <verb>
    → _MathNNLearning
         ├─ _MathNNExperience  (transitions, demos, prefer pairs, jobs)
         ├─ _RLAgentNet        (Q/RM/WM, Select/Observe, SupervisedQStep)
         ├─ _MathNNDevice      (train/infer placement)
         └─ _RLCheckpointMySQL (meta, transfer, quant, federate, promote)

Experience store

Buffer	Filled by
Transitions	Every RLAgentNet::Observe
Last state/action	Select + Observe
Demos	imitate start + Select, or load file
Prefer pairs	prefer good/bad and prefer pair
Jobs	nn learn jobs enqueue …

Verbs (implemented)

Verb	Real behavior
prefer	RM feedback on last state; auto pair buffer
prefer pair / train	Explicit pairs + batch RM train
curious	WM on + scale; warm on stored transitions
imitate start/stop/train	Demo record + BC via SupervisedQStep
distill	Weight copy + KD matching teacher Q
transfer	Full load; optional encoder fine-tune
adapt	RM plasticity + replay mix
quantize	Save + cal pass + NPU placement + JSON sidecar
shadow	Peer agent + dual-Q metrics on Select
curriculum	Phases + success-gated next
hierarchy	Options + HierarchyAct worker steps
whatif	Q on last state + WM-style horizon
align	RM + low curiosity + batch caution labels
federate	push/pull MySQL + avg ≈ KD toward global
promote	multi-metric gate (loss, pref, shadow)
scale	tiny/default/large/huge + device pref
pack	TutorOS, EmbodiedCraft, LifeCoach, Operator, MetaMind, WorldForge
jobs	enqueue / drain simple train jobs
experience	Store stats

Examples

[text]
nn learn imitate start life
# … agent acts …
nn learn imitate stop life
nn learn imitate train life 128

nn learn prefer good
nn learn prefer train life 32

nn learn distill life from life_focus_blocks 64
nn learn transfer minecraft space_engineers encoder
nn learn shadow minecraft on
nn learn promote minecraft multi
nn learn pack EmbodiedCraft
nn learn pack LifeCoach
nn learn whatif 0 3
nn learn federate life push
nn learn federate life avg

Files

• MathNNExperience.hpp/cpp
• MathNNLearning.hpp/cpp
• hooks in RLAgentNet.cpp (Observe → store, Select → demo/shadow)

Still future (honest)

• True ONNX NPU EP execution
• Cryptographic FedAvg / secure aggregation
• Full DAgger loop UI
• External async trainer process

---

Filed under: Uncategorized - @ July 18, 2026 6:42 am