Docs  /  Including the brainstorms with MathNN

Including the brainstorms with MathNN

This doc maps every major brainstorm from the recent stack onto MathNN, and shows how they share one access map (nn access).

Unified picture

                    L5  LLM coach / FT harvest / commander / ethics
                    L4  Minecraft body: DQN, unstuck, chat, MySQL, WM hub, perception
                    L3  Agent packs + RLAgentNet (minecraft dueling-dqn)
                    L2  nn learn + NeuralLearnMethods (Dyna/offline/DAgger/…)
                    L1  Simple multimodal encode (optional fuse into state)
                    L0  MathNN core (layers, optimizers, device)

MathNN is the plastic brain. Brainstorms supply regimes (packs), data (MySQL/chat/demos), imagination (WM), language (FT/LLM), and recovery (unstuck)—all of which should feed or gate MathNN, not replace it.

Brainstorm → MathNN inclusion

Brainstorm How it includes MathNN Access feature id Primary commands
Jump/place when stuck Forces practice transitions; rewards shape Q minecraft_unstuck unstuck on, pack apply corner_coach, method dagger
Epsilon / session packs Sets exploration for MathNN Select agent_packs pack apply default\|scout\|sniper\|eval_demo
Coach packs Human + shaping → RM / BC agent_pack_corner corner_coach, teach_live, reward / rlaif
Train/eval chat Visibility; optional FT text (not Q weights) minecraft_train_chat train-chat on, eval-chat on, speak
MySQL agent store Offline MathNN fuel (runs, feedback) minecraft_mysql pack mysql init, method offline
Infinity world model Curiosity + Dyna imagination into Q world_model_hub, nn_learn_dyna pack wm *, method dyna on
Neural learn methods Full train suite on agent minecraft nn_learn_methods nn learn method pack learn_all
Fine-tune LLM Language coach; RLAIF labels → RM llm_ft_harvest pack ft harvest/export, rlaif
Commander / EF plan Goals condition MathNN rewards minecraft_commander, executive_plan commander …, goal …
Minecraft × MathNN training Online step + nightly agent_minecraft, minecraft_dqn net on, train, nightly
Full stack One recipe brainstorm_stack nn access recipe brainstorm_stack

Data flows into MathNN

MC observe (88-D + grid/path/vision)
        │
        ▼
  Select: ε from pack ──► MathNN Q / legal mask (wall-guard)
        │
        ▼
  Act (client/bridge) + unstuck override (practice)
        │
        ▼
  Reward = env + stuck shaping + human + RM + curiosity
        │
        ▼
  Observe ──► TD train
        │         ├─ Dyna k imagined (WM)
        │         ├─ experience store (+ PER offline)
        │         └─ MySQL feedback/chat (async offline)
        ▼
  Language path: speak / !ai / FT harvest  (coach model, separate weights)

Division of labor (keep this clean)

Concern Owner Not
Motor Q / skills MathNN + RLAgentNet Free-form LLM
Jump/place recovery Unstuck + BC/Dyna Chat alone
Exploration schedule Packs (ε) Random CLI
Imagination WM + method dyna LLM fantasy rewards
Long-term memory of runs MySQL Process RAM only
“How do I use _AI?” LLM FT / docs Q-network
“Gather wood next” Commander → goal Direct key spam from LLM
Safety Shield / ethics / factions Hope

Operator entry points

Map everything

nn access map
nn access recipe brainstorm_stack
nn access probe brainstorm_stack
nn access probe world_model_hub
nn access probe nn_learn_methods
nn access probe agent_packs

minecraft bridge start
minecraft dqn pack stack enable train
minecraft dqn pack stack doctor
minecraft dqn pack synth run 32
# later:
nn learn method nightly minecraft 256
nn learn promote minecraft multi

See BrainstormStackComplete.md.

By goal

You want… Include
Better corners unstuck + corner_coach + learn_bc + prefer/rlaif
Sample efficiency net_curiosity + wm_dyna + method dyna + nightly
Multi-session memory mysql_full + offline + promote
Demo stream eval_demo + ε=0 (MathNN frozen)
Product Q&A ft_product harvest (LLM), not MathNN
Goal chaining commander + curriculum method pack

Access Hub features added

id Level
nn_learn_methods, nn_learn_dyna, nn_learn_offline, nn_learn_dagger L2
agent_packs, agent_pack_corner, agent_pack_net_curiosity L3
minecraft_unstuck, minecraft_train_chat, minecraft_mysql, world_model_hub L4
llm_ft_harvest, llm_ft_mc_coach, minecraft_commander, brainstorm_stack L5

Live probe hooks call AIAgentPacks / NeuralLearnMethods status where relevant.

Design rule

Every brainstorm either (a) changes MathNN weights (train), (b) changes MathNN inputs/masks (state, legal actions), (c) changes MathNN hyperparameters (ε, curiosity), or (d) runs a parallel language/memory system that may label rewards—but does not drive keys without a validated goal/action path.

That is how the “other few brainstorms” stay included with MathNN instead of as disconnected toys.

Generated from the project markdown docs on 2026-07-24. This is a static, self-contained site.