Experiment 03 — feed-forward network with plant-watering decision logic

Experiment date: July 31, 2026
Distribution: Sensory 0.1.1 for Windows

Question

Can a small Cognia feed-forward network turn plant and weather state into a practical recommendation while keeping learned estimation separate from hard safety rules?

This is not an agronomic model or real-world watering guidance. A synthetic teacher produces the dataset so the experiment remains local, reproducible, and easy to understand.

Input and output

The network receives nine values in this order:

1. soil moisture,
2. air temperature,
3. relative air humidity,
4. sunlight intensity,
5. rain probability,
6. hours since the last watering,
7. the cactus flag,
8. the herb flag,
9. the tomato flag.

The client and generator linearly map the first six physical values to [-1, 1]. Plant type is one-hot encoded. The three output neurons represent:

NO_WATER | LIGHT_WATER | DEEP_WATER

Architecture

9 sensors ──→ 9 fixed nonlinear features ──→ 3 decisions
     └──────────────── raw skip path ──────→┘

The fixed layer applies a feed-forward tanh transform and its weights are not trained. Softmax training updates only edges entering the final decision area. The skip path preserves raw information for the linear readout.

This limitation is deliberate and important: Cognia is not an autodiff framework, and the current runtime does not provide general backpropagation or deep credit assignment. The example therefore makes no end-to-end deep-training claim.

Synthetic teacher

The generator computes a continuous water-demand score. Soil dryness has the largest weight; heat, dry air, sunlight, and plant type make smaller contributions. Forecast rain reduces demand. Two thresholds map the score to three classes.

Generation uses seed 103 and rejection sampling until it has exactly 200 examples of each class. Balancing prevents the trivial strategy of always choosing the most common action. Time since watering is available as context, but the hard cooldown remains in the safety layer.

Post-network decision policy

The client converts the three scores to probabilities, selects their maximum, and then applies:

soil moisture >= 75%       → NO_WATER
time since watering < 6 h  → NO_WATER
highest confidence < 55%   → WAIT_AND_RECHECK
otherwise                  → network recommendation

This separates a statistical recommendation from protective invariants. A model error cannot bypass the ban on watering wet or just-watered soil.

Reproduction and measured result

The commands are in README-en.md. With the included dataset, seed 103, and 5000 iterations, this distribution build measured:

accuracy: 403/600 (67.2%)
MSE:      0.4355

This is training accuracy, not a generalization estimate. The result is reported without embellishment: the final local readout captures the main trend, but compound multilayer classification without deep credit assignment remains unsolved. A serious evaluation would generate a separate test set with another seed.

Verified illustrative scenarios after this training run:

What the experiment represents

• mapping familiar physical units into neural input;
• a real feed-forward layered path with no recurrence;
• a three-class softmax decision;
• REST deployment of the trained network;
• safe composition of a probabilistic model and deterministic policy;
• a transparent demonstration of Cognia's current deeper-learning limit.

Propsal

Proposal for Cognia 0.1.2