Skip to content

Latest commit

 

History

History
122 lines (88 loc) · 6.94 KB

BRAIN.md

File metadata and controls

122 lines (88 loc) · 6.94 KB

Neural Network Architecture

Primordium entities behave using a dynamic, graph-based neural network inspired by NEAT (NeuroEvolution of Augmenting Topologies).

From Matrix to Graph (Phase 28)

In earlier phases, entities used a fixed MLP (Multilayer Perceptron) architecture. Phase 28 introduces NEAT-lite, allowing the brain's topology to evolve over time alongside its weights. This enables the emergence of complex, specialized cognitive structures that are not limited by a fixed layer count or neuron density.

Topology

The brain starts with a standard minimal configuration but grows dynamically:

  • Initial Input Layer: 29 Neurons (14 Environment + 6 Recurrent + 9 Civilizational/Contextual)
  • Initial Hidden Layer: 6 Neurons
  • Initial Output Layer: 12 Neurons
  • Dynamic Growth: Through mutations, new hidden nodes and connections can be added indefinitely.
  • Node ID Mapping (Phase 63 Update):
    • Inputs: 0..28 (0-13 Env, 14-19 Memory, 20-28 Macro/Social)
    • Outputs: 29..40
    • Initial Hidden: 41..47

Inputs (Sensors)

Environmental Inputs (0-13)

ID Sensor Description
0 FoodDX X-distance to nearest food source
1 FoodDY Y-distance to nearest food source
2 Energy Internal energy level (0.0 to 1.0)
3 Density Local neighbor density
4 Phero Food pheromone strength at current location
5 Tribe Local tribe member count
6 KX Kin Centroid X (relative direction)
7 KY Kin Centroid Y (relative direction)
8 SA Signal A (Semantic Signal input)
9 SB Signal B (Semantic Signal input)
10 WL Wall/Boundary proximity
11 AG Age/Maturity status (0.0 to 1.0)
12 NT Nutrient Type of nearest food (0.0=Green, 1.0=Blue)
13 TP Internal Trophic Potential (0.0=Herbivore, 1.0=Carnivore)
ID Sensor Description
14-19 Memory Output values of the initial 6 hidden nodes from previous tick (T-1)
20 Hear Hearing Input (Sound intensity at location)
21 PartnerEnergy Energy level of potential mate
22 BuildPress Local demand for building structures
23 DigPress Local demand for digging canals
24 SharedGoal Lineage-wide goal reinforcement signal
25 SharedThreat Lineage-wide threat reinforcement signal
26 LinPop Macro population count of the lineage
27 LinEnergy Macro total energy consumption of the lineage
28 Overmind Active guidance signal from high-rank Alphas

Macro-Evolutionary Feedback (Phase 60)

The brain now incorporates Shared Memory inputs. When an entity dies to predation or finds massive food, it reinforces a global lineage signal. Every member of the lineage senses this via the SharedGoal and SharedThreat inputs, allowing for the emergence of "Instinctual Swarming" or "Collective Cowardice" without direct proximity.

Planetary Perception (Phase 62)

Advanced lineages perceive their global footprint. Inputs 26 and 27 (LinPop, LinEnergy) allow organisms to scale their aggression or reproductive investment based on their lineage's current dominance or vulnerability in the simulation.

Trophic Spectrum Influence (Phase 33)

With the introduction of the Trophic Potential (TP) gene, the brain must now navigate a sliding scale of dietary strategies rather than a binary choice.

  • Herbivore Dominance (TP < 0.5): The brain prioritizes NT inputs that match its metabolic_niche (usually Green). High sensitivity to Phero and FoodDX/Y is favored for efficient grazing.
  • Carnivore Dominance (TP > 0.5): The brain evolves to ignore Phero (food trails) in favor of Density and SA/SB signals from potential prey. Movement logic shifts from "seek resource" to "intercept target".
  • The Hybrid Trap: Entities with mid-range TP (0.4 - 0.6) must balance both strategies but suffer from reduced efficiency in both, often leading to specialized "opportunist" behaviors or extinction in highly competitive environments.

Outputs (Actions)

ID Action Threshold
29 MoveX Continuous (-1.0 to 1.0)
30 MoveY Continuous (-1.0 to 1.0)
31 Speed Continuous (Max speed modulation)
32 Aggro > 0.5 invokes attack state
33 Share > 0.5 transfers energy to tribe
34 Color Real-time color modulation (-1.0 to 1.0)
35 EmitSA > 0.5 emits Signal A
36 EmitSB > 0.5 emits Signal B
37 Bond > 0.5 initiates symbiosis/mating request
38 Dig > 0.5 convert Wall/Mountain -> Plains
39 Build > 0.5 convert Plains -> Wall/Nest/Outpost
40 Overmind > 0.5 emits lineage-wide Overmind signal (Alphas only)

Protected Cognitive Modules (Phase 62)

To prevent civilizational collapse in high-radiation eras, specialized castes develop Protected weight sets.

  • Soldiers: Weights leading to the Aggro output (32) are resistant to mutation.
  • Engineers: Weights leading to Dig and Build (38, 39) are resistant to mutation. This allows lineages to maintain their engineering and defense capabilities even when environmental stressors force rapid neural drift in other behavioral areas.

Topological Mutations

Evolution now acts on the structure of the brain through two primary mechanisms:

  1. Add Connection: A new connection is created between two previously unconnected nodes.
  2. Add Node: An existing connection is split. The old connection is disabled, and two new connections are created leading in and out of the new hidden node.

Innovation Tracking

To allow for successful crossover (sexual reproduction) between different topologies, every new structural mutation is assigned a global Innovation Number. During crossover, genes with matching innovation numbers are aligned, while disjoint/excess genes are inherited from the fitter parent, preventing the "competing conventions" problem.

Metabolic Cost of Complexity

Intelligence is not free. To prevent "bloat" (unnecessary complexity that doesn't provide a survival advantage), every structural element carries a metabolic maintenance cost added to the base idle metabolism:

  • Per Hidden Node: 0.02 energy / tick
  • Per Enabled Connection: 0.005 energy / tick

This creates a natural selection pressure for efficiency, where only complexity that significantly improves survival remains in the gene pool.

Neural Pruning (Phase 66.5)

To further combat topological bloat, a Neural Pruning mechanism has been introduced. Periodically, weak connections (weights below a configurable threshold) are removed from the brain's topology. This prevents "junk" connections from accumulating metabolic costs and degrading simulation performance during long-term evolutionary runs.

Activation Function

We use Tanh (Hyperbolic Tangent) for all nodes to allow negative values, mapping signals to [-1.0, 1.0].