Neural Network

Neural Network

The Neural Network in TheLifeOfPy determines how creatures process sensory input and choose actions. Each sensor and its connections function as a forest of trees, where each tree independently processes inputs.

The actuators with values equal to the highest activation in a tree are selected, and their corresponding actions are performed.

---

Forward Propagation Algorithm

Below is a simplified algorithm for processing sensory input and selecting actions:

Input:

• A list of input values corresponding to sensor nodes.
• Predefined sets of:
  • Nodes categorized as sensor, bias, hidden, and actuator.
  • Connections between nodes with associated weights and enabled/disabled states.
  • Sorted node list indicating the order of computation.
  • Grouped actuator nodes for activation resolution.

Algorithm:

1. Initialize Activations:

   • Set each sensor node’s activation to its corresponding input value.
   • Set each bias node’s activation to 1.
   • Set all other nodes’ activations to 0.

2. Propagate Activations:
   For each node in sorted order:

   • Skip if it is a sensor or bias node.
   • Initialize value as 0.
   • For each input connection to this node:
     • If the connection is enabled:
       • Add the product of the input node’s activation and the connection weight to value.
   • Store value as this node's activation.

3. Apply Activation Function (Winner-Takes-All):
   For each group of nodes:

   • Identify the actuator nodes in the group.
   • If none exist, skip.
   • Find the maximum activation value among them.
     • If the maximum value is negative:
       • Set all activations in the group to 0.
     • Otherwise:
       • Set the activation of nodes with the maximum value to 1, and all others to 0.

4. Return Activated Actuators:

   • Return the list of actuator nodes whose activation value is 1.

Sensors & Actuators

Sensors (Inputs)

Sensors provide environmental data to the neural network, influencing decision-making. If an actuator is marked (if visible), the relevant proximity sensor must be selected first, as it stores context for actuator calculations.

Random & Environmental Sensors

Code	Description	Output Range
RNs	Random noise generator	-1 (Low) → 1 (High)
FDi	Distance to nearest plant shrub	-1 (On it) → 1 (Farthest)
FAm	Density of plant shrubs nearby	-1 (None) → 1 (More than 10)

Proximity & Density Sensors

Code	Description	Output Range
ADi	Distance to nearest same-species critter	-1 (On it) → 1 (Farthest)
ODi	Distance to nearest other-species critter	-1 (On it) → 1 (Farthest)
CDi	Distance to nearest any-species critter	-1 (On it) → 1 (Farthest)
AAm	Density of same-species critters nearby	-1 (None) → 1 (More than 10)
OAm	Density of other-species critters nearby	-1 (None) → 1 (More than 10)
CAm	Density of any-species critters nearby	-1 (None) → 1 (More than 10)

Interaction & Context Sensors

Code	Description	Output Range
MsD	Distance to mouse pointer (if visible)	-1 (On it) → 1 (Farthest)
MSa	Whether mating signal was accepted	-1 (No) → 1 (Yes)

Self-State Sensors

Code	Description	Output Range
CEn	Current energy level	-1 (Empty) → 1 (Full)
CAg	Current age	-1 (Newborn) → 1 (Old)
CFi	Current fitness vs average	-1 (Low) → 1 (High)
RSt	Reproduction state	-1 (Not ready) → 0 (Ready) → 1 (Mating)
DSt	Defense state	-1 (Not activated) → 1 (Activated)

---

Actuators (Outputs)

Actuators determine how the critter responds to its environment. If an actuator is marked (if visible), the relevant proximity sensor must be selected first, as it stores context for actuator calculations.

Movement Actions

Code	Action
Mv	Move in a random direction
MvS	Move towards same-species critter (if visible)
MvO	Move towards other-species critter (if visible)
MvA	Move towards any-species critter (if visible)
MvF	Move towards nearest food source (if visible)
MvM	Move towards mouse pointer (if visible)

Defensive Actions

Code	Action
ADe	Activate defense mechanism
DDe	Deactivate defense mechanism

Avoidance Actions

Code	Action
AvS	Move away from same-species critter (if visible)
AvO	Move away from other-species critter (if visible)
AvA	Move away from any-species critter (if visible)
AvF	Move away from nearest food source (if visible)
AvM	Move away from mouse pointer (if visible)

Other Actions

Code	Action
Eat	Consume food if available
SMS	Send a mating signal to a nearby same-species critter (if visible)
Mte	Attempt to mate if a potential mate is found

---

Example Neural Networks

Example 1: Basic Food Detection

Description:
If food is nearby, the creature eats food; otherwise, it wanders.

Weights:

Sensor	Weight	Actuator
FDi	1	Mv
FDi	-1	Eat

---

Example 2: Interactive Feeding

Description:
If the mouse pointer is near, the creature moves towards it; if food is nearby, the creature eats the food.

Weights:

Sensor	Weight	Actuator
MsD	-1	MvM
FDi	-1	Eat

---

Example 3: Hostile interaction

Description:
If an organism of another species is nearby, the creature moves towards it and activates its defense mechanism. This behavior is designed for entities like Swordlings to attack other organisms. If no other species is near, it will wander and deactivate its defense mechanism.

Weights:

Sensor	Weight	Actuator
ODi	-1	MvO
ODi	-1	ADe
ODi	1	DDe
ODi	1	Mv