Reinforcement Learning in _AugmentedIntelligence

This is the operator hub for RL: what exists, how pieces fit together, and which docs/commands to use first.

Related maps:

Architecture (three layers)

┌─────────────────────────────────────────────────────────────┐
│  Typed commands / Lua / bridges (operator surface)          │
│  rl … · life rl … · minecraft dqn … · gaming rl … · nn …    │
└───────────────────────────┬─────────────────────────────────┘
                            │
┌───────────────────────────▼─────────────────────────────────┐
│  Domain agents (bespoke loops)                              │
│  MinecraftDQN · GamingRL · DrivingRL · Knowledge · LifeRL   │
│  + Executive Functions plans (high-level steps)             │
└───────────────────────────┬─────────────────────────────────┘
                            │
        ┌───────────────────┼───────────────────┐
        ▼                   ▼                   ▼
┌───────────────┐   ┌───────────────┐   ┌──────────────────┐
│ RLCore /      │   │ RLAgentNet    │   │ Remote / deep    │
│ Algorithms /  │   │ MathNN        │   │ trainer (optional)│
│ Curriculum    │   │ RLNeural      │   │ RLDistributed    │
│ smoke envs    │   │ local DQN…    │   │                  │
└───────────────┘   └───────────────┘   └──────────────────┘

1. Curriculum / smoke layer — safe synthetic RLEnvs; no live actuation.
2. Bespoke domain agents — real observations (Spigot, client, game mod, HUD).
3. MathNN / RLAgentNet — opt-in in-process neural policies (DQN, dueling, actor-critic, bandit, hierarchical, multi-head, world-model, reward-model).

Command families (quick map)

Full routing: rl dispatches through _RLFeatures (curriculum, real, net, cartpole, life). Prefer dry-run / smoke before live control.

In-process neural RL (MathNN)

Enable a neural policy on an agent:

nn demo
rl net modes
rl net enable minecraft
rl net mode minecraft dueling-dqn
rl net wm minecraft on
rl net curiosity minecraft 0.05
rl net reward minecraft good
rl net q minecraft
rl net loss minecraft
rl net save minecraft checkpoints/minecraft_mathnn.txt

Modes: dqn, dueling-dqn, actor-critic, bandit, hierarchical, multihead-dueling, worldmodel (dueling + curiosity), reward-model.

Auxiliary (combinable):

rl net rm <agent> on|off
rl net blend <agent> <alpha>      # α on env reward; (1−α) on reward model
rl net wm <agent> on|off
rl net curiosity <agent> <scale>

CartPole (ODE dynamics + MathNN):

rl cartpole 50
rl cartpole 50 dueling curiosity shield tripwire

Details: MathNN.md.

Domain agents

Minecraft DQN

• Observations: Spigot bridge (preferred) or RCON fallback; optional client keys.
• Neural: minecraft dqn net on / rl net enable minecraft.
• Goals: reach / mine / gather / earn.
• Executive Functions: high-level plan steps (wall-guard, net on, status) advance each train/eval step when enabled.

minecraft bridge start
minecraft dqn plan on
minecraft dqn goal mine iron_ore
minecraft dqn plan detail
minecraft dqn net on
minecraft dqn dry-run
minecraft dqn train 1000
minecraft dqn plan status

See MinecraftPluginTrainingGuide.md.

Gaming FPS / Strategy

gaming rl init
gaming rl dry-run on
gaming rl algorithm dqn
rl net enable gaming_fps
rl net mode gaming_fps actor-critic

See GamingRL.md, GameModBridges.md.

Driving

driving rl status
rl net enable driving
rl net mode driving dueling-dqn

See DrivingRL.md.

Life RL (advisory)

life rl domains
life rl smoke all 8
life rl net on
life rl suggest focus_blocks
life rl reward good

See LifeRL.md. No external side effects from the policy.

Knowledge / parenting / political

Specialized modules with stronger safety text. Prefer synthetic smoke and human review before anything that could affect real users.

Executive Functions + RL

Executive planning is not a second motor policy. It runs high-level setup/coaching steps (status, safety toggles, memory/awareness) while domain RL handles frame-to-frame actions.

Nested minecraft dqn train/step/observe from executive steps is blocked to avoid re-entrancy during auto-train.

Safe training ladder

1. Math smoke: nn demo, rl cartpole 50
2. Curriculum smoke: rl curriculum smoke all 8 96
3. Domain dry-run: observe / dry-run only (no live keys or RCON write)
4. Short train: small episode count; watch status/loss
5. Checkpoint: rl net save / domain save
6. Eval gate: promote only if metrics improve
7. Live: deployment stages via rl real stage …

Never jump from cold start to live desktop, medical, legal, finance, or public multiplayer bots.

RSI and Lua

RSI scripts get MathNN + executive Lua globals, for example:

ai_nn_modes()
ai_nn_policy_enable("life", "dueling-dqn", 16, 8)
rsi_sync_executive_plan(false)
rsi_advance_executive_plan(1)
executive_status()
ai_rl_status()

RSI cycles can seed and advance executive plans while the CommandRouter runs perception/reasoning commands. See RSI scripts under lua_scripts/RSI-*.lua.

File layout (code)

Troubleshooting

Version note

Documentation updated for the MathNN multi-learner registry, Life RL, Minecraft executive plan hooks, RSI executive integration, and unified rl command routing (2026).