{"id":2010,"date":"2026-07-18T06:41:31","date_gmt":"2026-07-18T13:41:31","guid":{"rendered":"http:\/\/macdaddy4sure.ai\/?p=2010"},"modified":"2026-07-18T06:41:31","modified_gmt":"2026-07-18T13:41:31","slug":"mathnn","status":"publish","type":"post","link":"http:\/\/macdaddy4sure.ai\/index.php\/2026\/07\/18\/mathnn\/","title":{"rendered":"MathNN"},"content":{"rendered":"\n

MathNN is the native C++ neural-network subsystem inside _AugmentedIntelligence.<\/p>\n\n\n\n

It is designed for small-to-medium in-process models where the application needs<\/p>\n\n\n\n

local learning without sending every update to TensorFlow, CUDA, or a remote<\/p>\n\n\n\n

policy server.<\/p>\n\n\n\n

The main current use case is reinforcement learning: MathNN can act as a local<\/p>\n\n\n\n

value network, policy network, or DQN Q-network for systems such as Minecraft,<\/p>\n\n\n\n

driving RL, gaming RL, strategy agents, Life RL, and math tutoring domains.<\/p>\n\n\n\n

RL hub:<\/strong> ReinforcementLearning.md (ReinforcementLearning.md) \u00b7 System map:<\/strong><\/p>\n\n\n\n

AugmentedIntelligenceSystems.md (AugmentedIntelligenceSystems.md) \u00b7 Life RL:<\/strong><\/p>\n\n\n\n

LifeRL.md (LifeRL.md)<\/p>\n\n\n\n

What MathNN Is<\/h2>\n\n\n\n

MathNN is a compact neural-network framework implemented in the repository’s C++<\/p>\n\n\n\n

codebase. The dense network core supports analytic backpropagation, optimizers,<\/p>\n\n\n\n

checkpoint save\/load, and command-line demos.<\/p>\n\n\n\n

MathNN is not meant to replace large model training stacks. Use CUDA, TensorFlow,<\/p>\n\n\n\n

TensorRT, ONNX Runtime, llama.cpp, or a remote policy server for large pretrained<\/p>\n\n\n\n

models, high-throughput vision models, LLMs, or heavy transformer training.<\/p>\n\n\n\n

Core Files<\/h2>\n\n\n\n
File<\/strong><\/td>Role<\/strong><\/td><\/tr>
MathNN.hpp \/ MathNN.cpp<\/td>Dense neural network core, activations, losses, optimizers, dropout, save\/load.<\/td><\/tr>
MathNNLayers.hpp \/ MathNNLayers.cpp<\/td>Conv2D, max pooling, layer norm, RNN, LSTM forward path, and attention forward path.<\/td><\/tr>
MathNNDevice.hpp \/ MathNNDevice.cpp<\/td>CPU\/GPU\/NPU\/APU\/Remote placement, probe, bench, agent overrides.<\/td><\/tr>
MathNNLearning.hpp \/ MathNNLearning.cpp<\/td>Domain learning functions: prefer, transfer, distill, scale, federate, \u2026<\/td><\/tr>
MathNNFeatures.hpp \/ MathNNFeatures.cpp<\/td>User-facing nn commands, Lua registration, and _NNPolicy.<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n

See also: MathNNLearningFunctions.md (MathNNLearningFunctions.md) \u00b7 MathDE.md (MathDE.md) \u00b7 BuildConfigurations.md (BuildConfigurations.md) \u00b7 RealityConstruct.md (RealityConstruct.md) \u00b7 SpaceEngineersRL.md (SpaceEngineersRL.md)<\/p>\n\n\n\n

RLNeural.hpp<\/strong> \/ <\/strong>RLNeural.cpp<\/strong><\/td>In-process learners: DQN, dueling DQN, actor-critic, bandit, hierarchical, multi-head dueling, reward model, world model.<\/strong><\/td><\/tr>
RLAgentNet.hpp \/ RLAgentNet.cpp<\/td>Per-agent neural policy registry with aux reward\/world models, checkpoint meta, Q\/loss\/feedback commands.<\/td><\/tr>
LifeRL.hpp \/ LifeRL.cpp<\/td>Advisory multi-domain “life” RL on MathNN (sleep, focus, chores, \u2026).<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n

What It Can Do<\/h2>\n\n\n\n

Dense Networks<\/h3>\n\n\n\n

The _NN class can build stacked dense networks:<\/p>\n\n\n\n

    \n
  • Add dense layers with configurable input\/output sizes.<\/li>\n\n\n\n
  • Choose layer activations.<\/li>\n\n\n\n
  • Run forward inference with Predict.<\/li>\n\n\n\n
  • Train with supervised TrainStep.<\/li>\n\n\n\n
  • Save and load model weights.<\/li>\n\n\n\n
  • Report parameter count and network summaries.<\/li>\n<\/ul>\n\n\n\n

    Supported activations:<\/p>\n\n\n\n

      \n
    • Linear<\/li>\n\n\n\n
    • ReLU<\/li>\n\n\n\n
    • LeakyReLU<\/li>\n\n\n\n
    • Sigmoid<\/li>\n\n\n\n
    • Tanh<\/li>\n\n\n\n
    • GELU<\/li>\n\n\n\n
    • Softmax<\/li>\n<\/ul>\n\n\n\n

      Supported losses:<\/p>\n\n\n\n

        \n
      • MSE<\/li>\n\n\n\n
      • CrossEntropy<\/li>\n\n\n\n
      • BCE<\/li>\n<\/ul>\n\n\n\n

        Supported optimizers:<\/p>\n\n\n\n

          \n
        • SGD<\/li>\n\n\n\n
        • Momentum<\/li>\n\n\n\n
        • Adam<\/li>\n<\/ul>\n\n\n\n

          Additional dense-network features:<\/p>\n\n\n\n

            \n
          • Inverted dropout during training.<\/li>\n\n\n\n
          • Fused stable gradients for softmax plus cross entropy.<\/li>\n\n\n\n
          • Fused stable gradients for sigmoid plus binary cross entropy.<\/li>\n\n\n\n
          • Xavier\/Glorot-style initialization for dense layers.<\/li>\n<\/ul>\n\n\n\n

            Architecture Layers<\/h3>\n\n\n\n

            MathNNLayers adds lower-level neural building blocks:<\/p>\n\n\n\n

            Component<\/strong><\/td>Capability<\/strong><\/td><\/tr>
            _Conv2D<\/td>Forward pass, backward pass, trainable weights, explicit channel-first image tensors.<\/td><\/tr>
            _MaxPool2D<\/td>Forward pass and backward pass with cached argmax locations.<\/td><\/tr>
            _LayerNorm<\/td>Forward pass, backward pass, trainable gamma\/beta.<\/td><\/tr>
            _RNNCell<\/td>Elman RNN forward steps plus sequence training with backprop-through-time.<\/td><\/tr>
            _LSTMCell<\/td>Forward\/inference path. Local BPTT training is not exposed; route larger LSTM training to a remote trainer.<\/td><\/tr>
            _Attention<\/td>Scaled dot-product attention and multi-head attention forward\/inference, with optional causal masking.<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n

            Tensor convention for image-like layers is flat row-major storage with explicit<\/p>\n\n\n\n

            shapes. Images use channels-first layout: C * H * W.<\/p>\n\n\n\n

            RL Policy Adapter<\/h3>\n\n\n\n

            _NNPolicy wraps _NN as a value or policy network:<\/p>\n\n\n\n

              \n
            • ActionValues(state) returns Q-values for all actions.<\/li>\n\n\n\n
            • GreedyAction(state) picks the highest-valued action.<\/li>\n\n\n\n
            • UpdateQ(state, action, target) trains one selected action toward a target.<\/li>\n\n\n\n
            • Save(path) and Load(path) persist the policy.<\/li>\n<\/ul>\n\n\n\n

              The default policy shape is:<\/p>\n\n\n\n

              [text]
              <\/em>state_dim -> hidden ReLU -> hidden ReLU -> action_dim Linear<\/p>\n\n\n\n

              This is enough for many DQN-style value functions where the network predicts one<\/p>\n\n\n\n

              Q-value per action.<\/p>\n\n\n\n

              In-Process DQN<\/h3>\n\n\n\n

              _NNQLearner builds a local DQN around _NNPolicy:<\/p>\n\n\n\n

                \n
              • Online network.<\/li>\n\n\n\n
              • Target network.<\/li>\n\n\n\n
              • Replay buffer.<\/li>\n\n\n\n
              • Epsilon-greedy action selection.<\/li>\n\n\n\n
              • TD target training: reward + gamma * max(Q(next_state)).<\/li>\n\n\n\n
              • Target network sync.<\/li>\n\n\n\n
              • Checkpoint save\/load.<\/li>\n<\/ul>\n\n\n\n

                _RLNeural::TrainDQN provides a training loop over any RLEnv implementation.<\/p>\n\n\n\n

                That makes MathNN reusable across game, robotics, driving, and knowledge<\/p>\n\n\n\n

                environments when those environments expose state, action, reward, and terminal<\/p>\n\n\n\n

                signals.<\/p>\n\n\n\n

                Per-Agent Neural Registry<\/h3>\n\n\n\n

                _RLAgentNet lets individual agents opt into a MathNN-backed learner without<\/p>\n\n\n\n

                rewriting their existing heuristic or tabular paths.<\/p>\n\n\n\n

                Primary modes:<\/strong> dqn, dueling-dqn, actor-critic, bandit,<\/p>\n\n\n\n

                hierarchical, multihead-dueling, worldmodel (dueling + curiosity),<\/p>\n\n\n\n

                reward-model (bandit + preference head).<\/p>\n\n\n\n

                Auxiliary (combinable):<\/strong><\/p>\n\n\n\n

                  \n
                • Reward model: rl net rm <agent> on, rl net reward <agent> good|bad, rl net blend <agent> 0.7<\/li>\n\n\n\n
                • World model \/ curiosity: rl net wm <agent> on, rl net curiosity <agent> 0.05<\/li>\n<\/ul>\n\n\n\n

                  Supported agent examples:<\/p>\n\n\n\n

                    \n
                  • minecraft, driving, gaming_fps, strategy<\/li>\n\n\n\n
                  • life, life_sleep_routine, life_focus_blocks, \u2026<\/li>\n\n\n\n
                  • math tutoring domains, human, cybernetic, desktop_support<\/li>\n<\/ul>\n\n\n\n

                    Checkpoints write a .meta sidecar (mode, dims) plus optional .rm \/ .wm files.<\/p>\n\n\n\n

                    Life RL<\/h3>\n\n\n\n

                    Advisory multi-domain daily-living RL (LifeRL). Synthetic smoke by default;<\/p>\n\n\n\n

                    sensitive domains (mood, budget, social) are bandit\/preference oriented and never<\/p>\n\n\n\n

                    auto-execute external side effects. See docs\/LifeRL.md.<\/p>\n\n\n\n

                    Commands<\/h2>\n\n\n\n

                    Mathematics RL Domains (MathNN)<\/h3>\n\n\n\n

                    Six STEM tutoring domains train MathNN DQN policies over safe synthetic<\/p>\n\n\n\n

                    curriculum environments (category mathematics). Each domain maps tutoring<\/p>\n\n\n\n

                    actions (simplify, apply theorem, hypothesis test, etc.) into the shared<\/p>\n\n\n\n

                    eight-value observation used by the RL curriculum smoke envs.<\/p>\n\n\n\n

                    Domain ID<\/strong><\/td>Subject<\/strong><\/td><\/tr>
                    math_algebra<\/td>Algebra<\/td><\/tr>
                    math_geometry<\/td>Geometry<\/td><\/tr>
                    math_trigonometry<\/td>Trigonometry<\/td><\/tr>
                    math_probability_statistics<\/td>Probability and Statistics<\/td><\/tr>
                    math_linear_algebra<\/td>Linear Algebra<\/td><\/tr>
                    math_finite_math<\/td>Finite Math<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n

                    [text]
                    <\/em>nn math domains
                    nn math domain math_algebra
                    nn math smoke mathematics 8 72
                    nn math train math_algebra 200
                    nn math net enable math_algebra
                    rl net save math_algebra checkpoints\/math_algebra_mathnn.txt
                    rl curriculum domains mathematics
                    rl curriculum smoke math_trigonometry 8 78<\/p>\n\n\n\n

                    Aliases: algebra, geometry, trigonometry, probability_statistics,<\/p>\n\n\n\n

                    linear_algebra, finite_math, and mathematics (smoke all six).<\/p>\n\n\n\n

                    MathNN Demos<\/h3>\n\n\n\n

                    Run these in the _AugmentedIntelligence command interface:<\/p>\n\n\n\n

                    [text]
                    <\/em>nn status
                    nn demo
                    nn conv-demo
                    nn rnn-demo
                    nn attn-demo<\/p>\n\n\n\n

                    Command behavior:<\/p>\n\n\n\n

                    Command<\/strong><\/td>What it tests<\/strong><\/td><\/tr>
                    nn status<\/td>Confirms the MathNN command surface is registered.<\/td><\/tr>
                    nn demo<\/td>Trains a small dense network on XOR with BCE and Adam. This is the main smoke test for dense forward\/backprop\/training.<\/td><\/tr>
                    nn conv-demo<\/td>Runs Conv2D and MaxPool over an 8×8 input and prints output shapes.<\/td><\/tr>
                    nn rnn-demo<\/td>Trains a small RNN echo task and prints loss movement.<\/td><\/tr>
                    nn attn-demo<\/td>Runs multi-head causal attention over a small sequence.<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n

                    Lua surface:<\/p>\n\n\n\n

                    [text]
                    <\/em>ai_nn_xor_demo()<\/p>\n\n\n\n

                    This runs the XOR demo and returns the final loss.<\/p>\n\n\n\n

                    RLAgentNet Commands<\/h3>\n\n\n\n

                    [text]
                    <\/em>rl net status
                    rl net modes
                    rl net enable <agent>
                    rl net disable <agent>
                    rl net mode <agent> <dqn|dueling-dqn|actor-critic|bandit|hierarchical|multihead-dueling|worldmodel|reward-model>
                    rl net save <agent> <path>
                    rl net load <agent> <path>
                    rl net q <agent> [top_k]
                    rl net loss <agent>
                    rl net reward <agent> good|bad|<n>
                    rl net blend <agent> <alpha>
                    rl net curiosity <agent> <scale>
                    rl net wm <agent> on|off
                    rl net rm <agent> on|off<\/p>\n\n\n\n

                    Examples:<\/p>\n\n\n\n

                    [text]
                    <\/em>rl net enable driving
                    rl net mode gaming_fps actor-critic
                    rl net wm minecraft on
                    rl net curiosity minecraft 0.05
                    rl net reward gaming_fps good
                    rl net save driving checkpoints\/driving_mathnn.txt
                    rl net load driving checkpoints\/driving_mathnn.txt
                    rl cartpole 50 dueling curiosity
                    life rl smoke sleep_routine 12
                    life rl suggest focus_blocks<\/p>\n\n\n\n

                    Minecraft MathNN Commands<\/h3>\n\n\n\n

                    The Minecraft DQN subsystem can use MathNN through its neural-net mode:<\/p>\n\n\n\n

                    [text]
                    <\/em>minecraft dqn net on
                    minecraft dqn net off
                    minecraft dqn net status
                    minecraft dqn net save <path>
                    minecraft dqn net load <path><\/p>\n\n\n\n

                    Executive Functions (high-level setup steps each train\/eval tick):<\/p>\n\n\n\n

                    [text]
                    <\/em>minecraft dqn plan on
                    minecraft dqn plan detail
                    minecraft dqn plan next 2
                    minecraft dqn plan status<\/p>\n\n\n\n

                    Typical Minecraft startup sequence:<\/p>\n\n\n\n

                    [text]
                    <\/em>minecraft dqn observe
                    minecraft dqn dry-run
                    minecraft dqn plan on
                    minecraft dqn goal mine iron_ore
                    minecraft dqn net on
                    minecraft dqn net status
                    minecraft dqn train 1000<\/p>\n\n\n\n

                    Use short training runs first. Stop and inspect behavior before letting it run<\/p>\n\n\n\n

                    for long periods:<\/p>\n\n\n\n

                    [text]
                    <\/em>minecraft dqn off
                    minecraft dqn status
                    minecraft dqn net save minecraft_mathnn_checkpoint.txt<\/p>\n\n\n\n

                    How MathNN Fits Into RL<\/h2>\n\n\n\n

                    A MathNN-backed RL agent uses this loop:<\/p>\n\n\n\n

                      \n
                    1. The domain converts the current world into a numeric state vector.<\/li>\n\n\n\n
                    2. MathNN predicts Q-values for available actions.<\/li>\n\n\n\n
                    3. The agent chooses an action, often epsilon-greedy during training.<\/li>\n\n\n\n
                    4. The environment executes the action.<\/li>\n\n\n\n
                    5. The domain computes reward, next state, and terminal status.<\/li>\n\n\n\n
                    6. _NNQLearner stores the transition in replay memory.<\/li>\n\n\n\n
                    7. Training samples a minibatch and updates the online network.<\/li>\n\n\n\n
                    8. The target network is periodically synchronized.<\/li>\n\n\n\n
                    9. A checkpoint is saved after useful behavior appears.<\/li>\n<\/ol>\n\n\n\n

                      The most important part is the state\/reward\/action design. MathNN can only learn<\/p>\n\n\n\n

                      from the signals it is given. If the state vector does not show walls, velocity,<\/p>\n\n\n\n

                      health, objective distance, inventory, or other useful facts, the network cannot<\/p>\n\n\n\n

                      learn those facts. If the reward function pays the wrong behavior, the network<\/p>\n\n\n\n

                      will optimize the wrong behavior.<\/p>\n\n\n\n

                      Current Capability Matrix<\/h2>\n\n\n\n
                      Area<\/strong><\/td>Status<\/strong><\/td>Notes<\/strong><\/td><\/tr>
                      Dense MLP inference<\/td>Available<\/td>_NN::Predict.<\/td><\/tr>
                      Dense MLP training<\/td>Available<\/td>_NN::TrainStep.<\/td><\/tr>
                      Activations<\/td>Available<\/td>Linear, ReLU, LeakyReLU, Sigmoid, Tanh, GELU, Softmax.<\/td><\/tr>
                      Losses<\/td>Available<\/td>MSE, cross entropy, BCE.<\/td><\/tr>
                      Optimizers<\/td>Available<\/td>SGD, Momentum, Adam.<\/td><\/tr>
                      Dropout<\/td>Available<\/td>Training-time inverted dropout.<\/td><\/tr>
                      Save\/load<\/td>Available<\/td>Native MathNN checkpoint format.<\/td><\/tr>
                      Conv2D<\/td>Available<\/td>Forward\/backward implemented; gradient-check before production training.<\/td><\/tr>
                      MaxPool2D<\/td>Available<\/td>Forward\/backward implemented.<\/td><\/tr>
                      LayerNorm<\/td>Available<\/td>Forward\/backward implemented.<\/td><\/tr>
                      RNN<\/td>Available<\/td>Elman RNN with BPTT sequence training.<\/td><\/tr>
                      LSTM<\/td>Forward only<\/td>Use remote trainer for full local sequence training needs.<\/td><\/tr>
                      Attention<\/td>Forward only<\/td>Scaled dot-product and multi-head inference helpers.<\/td><\/tr>
                      DQN<\/td>Available<\/td>_NNQLearner with online\/target nets and replay.<\/td><\/tr>
                      Dueling DQN<\/td>Available<\/td>_NNDuelingQLearner V+A.<\/td><\/tr>
                      Actor-critic<\/td>Available<\/td>_NNActorCriticLearner policy+value.<\/td><\/tr>
                      Bandit<\/td>Available<\/td>_NNBanditLearner.<\/td><\/tr>
                      Hierarchical RL<\/td>Available<\/td>skill manager + worker DQN.<\/td><\/tr>
                      Multi-head dueling<\/td>Available<\/td>skill-group advantages.<\/td><\/tr>
                      Reward model<\/td>Available<\/td>preference BCE + blend into Observe.<\/td><\/tr>
                      World model<\/td>Available<\/td>\u0394s+reward predict + curiosity bonus.<\/td><\/tr>
                      Life RL<\/td>Available<\/td>advisory multi-domain (life rl \u2026).<\/td><\/tr>
                      Per-agent neural policies<\/td>Available<\/td>RLAgentNet opt-in registry.<\/td><\/tr>
                      LLM behavior<\/td>Not MathNN<\/td>Use the LLM subsystem; MathNN can score states\/actions, not generate rich language.<\/td><\/tr>
                      Large CUDA training<\/td>Not MathNN<\/td>Use CUDA\/TensorFlow\/remote training for large models.<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n

                      What MathNN Is Good For<\/h2>\n\n\n\n

                      Use MathNN for:<\/p>\n\n\n\n

                        \n
                      • Local DQN value functions.<\/li>\n\n\n\n
                      • Small supervised models.<\/li>\n\n\n\n
                      • RL action scoring.<\/li>\n\n\n\n
                      • Agent-specific policies.<\/li>\n\n\n\n
                      • Lightweight experiments that should compile into _AugmentedIntelligence.<\/li>\n\n\n\n
                      • Fast smoke tests for learning loops without external services.<\/li>\n\n\n\n
                      • Simple RNN sequence experiments.<\/li>\n\n\n\n
                      • Small perception experiments using Conv2D and pooling.<\/li>\n<\/ul>\n\n\n\n

                        Use another subsystem for:<\/p>\n\n\n\n

                          \n
                        • LLM chat and reasoning.<\/li>\n\n\n\n
                        • Image generation.<\/li>\n\n\n\n
                        • Large CNNs or transformers.<\/li>\n\n\n\n
                        • Production-scale CUDA training.<\/li>\n\n\n\n
                        • Model import\/export pipelines such as ONNX.<\/li>\n\n\n\n
                        • High-volume offline training jobs.<\/li>\n<\/ul>\n\n\n\n

                          Practical Training Guidance<\/h2>\n\n\n\n

                          Start with the demos before training a domain:<\/p>\n\n\n\n

                          [text]
                          <\/em>nn status
                          nn demo
                          nn conv-demo
                          nn rnn-demo
                          nn attn-demo<\/p>\n\n\n\n

                          Then validate the target domain without learning:<\/p>\n\n\n\n

                          [text]
                          <\/em>minecraft dqn observe
                          minecraft dqn dry-run<\/p>\n\n\n\n

                          For any RL domain, verify:<\/p>\n\n\n\n

                            \n
                          • State values change when the world changes.<\/li>\n\n\n\n
                          • Legal actions match what the agent can actually do.<\/li>\n\n\n\n
                          • Rewards increase for intended progress.<\/li>\n\n\n\n
                          • Terminal events are marked correctly.<\/li>\n\n\n\n
                          • Safety shields and stop commands work.<\/li>\n<\/ul>\n\n\n\n

                            Only after that should you enable MathNN training:<\/p>\n\n\n\n

                            [text]
                            <\/em>rl net enable <agent><\/p>\n\n\n\n

                            or for Minecraft:<\/p>\n\n\n\n

                            [text]
                            <\/em>minecraft dqn net on
                            minecraft dqn train 1000<\/p>\n\n\n\n

                            Save checkpoints when behavior improves:<\/p>\n\n\n\n

                            [text]
                            <\/em>rl net save <agent> checkpoints\/<agent>_mathnn.txt
                            minecraft dqn net save minecraft_mathnn_checkpoint.txt<\/p>\n\n\n\n

                            Troubleshooting<\/h2>\n\n\n\n
                            Problem<\/strong><\/td>Likely cause<\/strong><\/td>What to check<\/strong><\/td><\/tr>
                            nn status is not recognized<\/td>Command dispatcher is not routing nn commands.<\/td>Check _MathNNFeatures::HandleSpeechCommand registration.<\/td><\/tr>
                            nn demo loss does not decrease<\/td>Training path, optimizer, or numeric issue.<\/td>Rebuild, rerun, and compare XOR predictions.<\/td><\/tr>
                            RL agent never improves<\/td>Bad state\/reward\/action wiring.<\/td>Inspect observations, rewards, terminal flags, and action execution before tuning the network.<\/td><\/tr>
                            Minecraft walks into walls<\/td>State\/action safety issue, not just a network issue.<\/td>Use minecraft dqn observe, minecraft dqn wall-guard on, and staged arenas.<\/td><\/tr>
                            Checkpoint load fails<\/td>Wrong path, incompatible model shape, or inactive agent.<\/td>Enable\/configure the same agent and use a matching checkpoint.<\/td><\/tr>
                            Training is unstable<\/td>Learning rate, reward scale, or exploration is too aggressive.<\/td>Lower learning rate, normalize rewards, reduce action space, and train in shorter sessions.<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n

                            Safety Notes<\/h2>\n\n\n\n

                            MathNN can select actions automatically when connected to an RL domain. Treat<\/p>\n\n\n\n

                            live-control domains differently from simulations:<\/p>\n\n\n\n

                              \n
                            • Prefer simulation or dry-run first.<\/li>\n\n\n\n
                            • Keep explicit stop commands available.<\/li>\n\n\n\n
                            • Use safety shields for driving, robotics, Minecraft, and desktop control.<\/li>\n\n\n\n
                            • Keep high-stakes domains in advisory, simulated, or human-approved modes.<\/li>\n\n\n\n
                            • Do not deploy learned behavior directly into real-world control without<\/li>\n<\/ul>\n\n\n\n

                                independent validation.<\/p>\n\n\n\n

                              Short Summary<\/h2>\n\n\n\n

                              MathNN gives _AugmentedIntelligence a local neural-learning layer. It can train<\/p>\n\n\n\n

                              small dense networks, run several neural layer demos, act as a DQN value network,<\/p>\n\n\n\n

                              and provide opt-in neural policies for RL agents. Its strongest current role is<\/p>\n\n\n\n

                              local reinforcement learning support, especially when paired with carefully<\/p>\n\n\n\n

                              designed states, rewards, legal actions, and safety guards.<\/p>\n\n\n\n

                              Multi-device placement (CPU \/ GPU \/ NPU \/ APU)<\/h2>\n\n\n\n

                              `             ext<\/p>\n\n\n\n

                              nn device status<\/p>\n\n\n\n

                              nn device list<\/p>\n\n\n\n

                              nn device refresh<\/p>\n\n\n\n

                              nn device bench all 256<\/p>\n\n\n\n

                              nn device set global train gpu<\/p>\n\n\n\n

                              nn device set global infer npu<\/p>\n\n\n\n

                              nn device set global dtype fp16<\/p>\n\n\n\n

                              nn device set agent minecraft train gpu infer gpu<\/p>\n\n\n\n

                              nn device set agent life infer npu<\/p>\n\n\n\n

                              nn device resolve space_engineers<\/p>\n\n\n\n

                              nn device clear agent life<\/p>\n\n\n\n

                              `<\/p>\n\n\n\n

                                \n
                              • CPU<\/strong> \ufffd MatrixAcceleration AVX\/AVX-512 GEMM (default, correctness).<\/li>\n\n\n\n
                              • GPU<\/strong> \ufffd CUDA probe + hybrid staged buffers (host GEMM until cuBLAS kernel lands).<\/li>\n\n\n\n
                              • NPU<\/strong> \ufffd env hints (OPENVINO_DEVICE, QNN_SDK_ROOT, DML_VISIBLE_DEVICES); placeholder GEMM + ONNX EP path later.<\/li>\n\n\n\n
                              • APU<\/strong> \ufffd hybrid policy (CPU train path + optional iGPU\/NPU prefs).<\/li>\n\n\n\n
                              • Remote<\/strong> \ufffd AI_REMOTE_CUDA_HOST for HardwareAcceleration workers.<\/li>\n<\/ul>\n\n\n\n

                                Agent status (nn agent status <id>) includes device_train \/ device_infer resolved devices.<\/p>\n","protected":false},"excerpt":{"rendered":"

                                MathNN is the native C++ neural-network subsystem inside _AugmentedIntelligence. It is designed for small-to-medium in-process models where the application needs local learning without sending every update to TensorFlow, CUDA, or a remote policy server. The main current use case is reinforcement learning: MathNN can act as a local value network, policy network, or DQN Q-network […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-2010","post","type-post","status-publish","format-standard","hentry","category-uncategorized"],"_links":{"self":[{"href":"http:\/\/macdaddy4sure.ai\/index.php\/wp-json\/wp\/v2\/posts\/2010","targetHints":{"allow":["GET"]}}],"collection":[{"href":"http:\/\/macdaddy4sure.ai\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/macdaddy4sure.ai\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/macdaddy4sure.ai\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/macdaddy4sure.ai\/index.php\/wp-json\/wp\/v2\/comments?post=2010"}],"version-history":[{"count":1,"href":"http:\/\/macdaddy4sure.ai\/index.php\/wp-json\/wp\/v2\/posts\/2010\/revisions"}],"predecessor-version":[{"id":2011,"href":"http:\/\/macdaddy4sure.ai\/index.php\/wp-json\/wp\/v2\/posts\/2010\/revisions\/2011"}],"wp:attachment":[{"href":"http:\/\/macdaddy4sure.ai\/index.php\/wp-json\/wp\/v2\/media?parent=2010"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/macdaddy4sure.ai\/index.php\/wp-json\/wp\/v2\/categories?post=2010"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/macdaddy4sure.ai\/index.php\/wp-json\/wp\/v2\/tags?post=2010"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}