neat / neat\genome.py
eyad-silx's picture
Upload neat\genome.py with huggingface_hub
8e3abc5 verified
"""NEAT Genome implementation.
This module implements the core NEAT genome structure and operations.
Each genome represents a neural network with nodes (neurons) and connections (synapses).
The genome can be mutated to evolve the network structure and weights over time.
"""
from dataclasses import dataclass
import jax.numpy as jnp
import jax.random as jrandom
from typing import Dict, List, Tuple, Optional
import time
import random
import numpy as np
@dataclass
class NodeGene:
"""Node gene containing activation function and type.
Attributes:
node_id: Unique identifier for this node
node_type: Type of node ('input', 'hidden', 'recurrent', or 'output')
activation: Activation function ('tanh', 'relu', 'sigmoid', or 'linear')
"""
node_id: int
node_type: str # 'input', 'hidden', 'recurrent', or 'output'
activation: str # 'tanh', 'relu', 'sigmoid', or 'linear'
@dataclass
class ConnectionGene:
"""Connection gene containing connection properties.
Attributes:
source: ID of source node
target: ID of target node
weight: Connection weight
enabled: Whether connection is enabled
innovation: Unique innovation number for this connection
"""
source: int
target: int
weight: float
enabled: bool = True
innovation: int = 0
class Genome:
"""NEAT Genome implementation.
A genome represents a neural network as a collection of node and connection genes.
The network topology can be modified through mutation operations.
Attributes:
input_size: Number of input nodes
output_size: Number of output nodes
node_genes: Dictionary mapping node IDs to NodeGene objects
connection_genes: List of ConnectionGene objects
key: Random key for reproducible randomness
innovation_number: Counter for assigning unique innovation numbers
"""
def __init__(self, input_size: int, output_size: int):
"""Initialize genome with specified number of inputs and outputs.
Args:
input_size: Number of input nodes
output_size: Number of output nodes (must be 3 for volleyball)
"""
self.input_size = input_size
self.output_size = output_size
self.node_genes: Dict[int, NodeGene] = {}
self.connection_genes: List[ConnectionGene] = []
# Initialize random key
timestamp = int(time.time() * 1000)
self.key = jrandom.PRNGKey(hash((input_size, output_size, timestamp)) % (2**32))
# Counter for assigning unique innovation numbers
self.innovation_number = 0
# Initialize minimal network structure
self._init_minimal()
def _init_minimal(self):
"""Initialize minimal feed-forward network structure.
Network structure:
- Input nodes [0-7]: Game state inputs
- Hidden layer 1 [8-15]: First processing layer (8 nodes)
- Hidden layer 2 [16-23]: Second processing layer (8 nodes)
- Output nodes [24-26]: Action outputs (left, right, jump)
Using larger initial weights for faster learning:
- Input->Hidden1: N(0, 2.0) for strong initial responses
- Hidden1->Hidden2: N(0, 2.0) for feature processing
- Hidden2->Output: N(0, 4.0) for decisive actions
"""
# Create input nodes (0-7)
for i in range(8): # Only 8 inputs used
self.node_genes[i] = NodeGene(
node_id=i,
node_type='input',
activation='linear' # Input nodes are always linear
)
# Create first hidden layer (8-15)
hidden1_size = 8
hidden1_start = 8 # Right after inputs
for i in range(hidden1_size):
node_id = hidden1_start + i
self.node_genes[node_id] = NodeGene(
node_id=node_id,
node_type='hidden',
activation='relu' # ReLU for faster learning
)
# Connect all inputs to this hidden node
for input_id in range(8):
weight = float(jrandom.normal(self.key) * 2.0)
self.connection_genes.append(ConnectionGene(
source=input_id,
target=node_id,
weight=weight,
enabled=True,
innovation=self.innovation_number
))
self.innovation_number += 1
# Create second hidden layer (16-23)
hidden2_size = 8
hidden2_start = hidden1_start + hidden1_size
for i in range(hidden2_size):
node_id = hidden2_start + i
self.node_genes[node_id] = NodeGene(
node_id=node_id,
node_type='hidden',
activation='relu' # ReLU for faster learning
)
# Connect all hidden1 nodes to this hidden2 node
for h1_id in range(hidden1_start, hidden1_start + hidden1_size):
weight = float(jrandom.normal(self.key) * 2.0)
self.connection_genes.append(ConnectionGene(
source=h1_id,
target=node_id,
weight=weight,
enabled=True,
innovation=self.innovation_number
))
self.innovation_number += 1
# Create output nodes (24-26)
output_start = hidden2_start + hidden2_size
for i in range(self.output_size):
node_id = output_start + i
self.node_genes[node_id] = NodeGene(
node_id=node_id,
node_type='output',
activation='tanh' # tanh for [-1,1] outputs
)
# Connect all hidden2 nodes to this output
for h2_id in range(hidden2_start, hidden2_start + hidden2_size):
weight = float(jrandom.normal(self.key) * 4.0) # Larger weights for outputs
self.connection_genes.append(ConnectionGene(
source=h2_id,
target=node_id,
weight=weight,
enabled=True,
innovation=self.innovation_number
))
self.innovation_number += 1
def mutate(self, config: Dict):
"""Mutate the genome by modifying weights and network structure.
Args:
config: Dictionary containing mutation parameters:
- weight_mutation_rate: Probability of mutating each weight
- weight_mutation_power: Standard deviation for weight mutations
- add_node_rate: Probability of adding a new node
- add_connection_rate: Probability of adding a new connection
"""
# Mutate connection weights
for conn in self.connection_genes:
if jrandom.uniform(self.key) < config['weight_mutation_rate']:
# Get new random key
self.key, subkey = jrandom.split(self.key)
# Add random value from normal distribution
conn.weight += float(jrandom.normal(subkey) * config['weight_mutation_power'])
# Add new nodes (disabled for now since we're using fixed topology)
if config['add_node_rate'] > 0:
if jrandom.uniform(self.key) < config['add_node_rate']:
self._add_node()
# Add new connections (disabled for now)
if config['add_connection_rate'] > 0:
if jrandom.uniform(self.key) < config['add_connection_rate']:
self._add_connection()
def _add_node(self):
"""Add a new node by splitting an existing connection."""
if not self.connection_genes:
return
# Choose a random connection to split
conn_to_split = np.random.choice(self.connection_genes)
conn_to_split.enabled = False
# Create new node
new_node_id = max(self.node_genes.keys()) + 1
self.node_genes[new_node_id] = NodeGene(
node_id=new_node_id,
node_type='hidden',
activation='relu'
)
# Create two new connections
self.connection_genes.extend([
ConnectionGene(
source=conn_to_split.source,
target=new_node_id,
weight=1.0,
enabled=True,
innovation=self.innovation_number
),
ConnectionGene(
source=new_node_id,
target=conn_to_split.target,
weight=conn_to_split.weight,
enabled=True,
innovation=self.innovation_number + 1
)
])
self.innovation_number += 2
def _add_connection(self):
"""Add a new connection between two unconnected nodes."""
# Get list of all possible connections
existing_connections = {(c.source, c.target) for c in self.connection_genes}
possible_connections = []
for source in self.node_genes:
for target in self.node_genes:
# Skip if connection already exists
if (source, target) in existing_connections:
continue
# Skip if would create cycle (except recurrent)
if self.node_genes[source].node_type != 'recurrent' and \
self.would_create_cycle(source, target):
continue
possible_connections.append((source, target))
if possible_connections:
# Choose random connection
source, target = random.choice(possible_connections)
# Create new connection
weight = float(jrandom.normal(self.key) * 1.0)
self.connection_genes.append(ConnectionGene(
source=source,
target=target,
weight=weight,
enabled=True,
innovation=self.innovation_number
))
self.innovation_number += 1
def would_create_cycle(self, source: int, target: int) -> bool:
"""Check if adding connection would create cycle in network.
Args:
source: Source node ID
target: Target node ID
Returns:
True if connection would create cycle, False otherwise
"""
# Skip cycle detection for recurrent connections
if self.node_genes[source].node_type == 'recurrent' or \
self.node_genes[target].node_type == 'recurrent':
return False
# Do depth-first search from target to see if we can reach source
visited = set()
def dfs(node: int) -> bool:
if node == source:
return True
if node in visited:
return False
visited.add(node)
for conn in self.connection_genes:
if conn.source == node and conn.enabled:
if dfs(conn.target):
return True
return False
return dfs(target)
def add_node_between(self, source: int, target: int):
"""Add a new node between two nodes, splitting an existing connection.
Args:
source: Source node ID
target: Target node ID
"""
# Find and disable the existing connection
for conn in self.connection_genes:
if conn.source == source and conn.target == target and conn.enabled:
conn.enabled = False
# Create new node
new_id = max(self.node_genes.keys()) + 1
self.node_genes[new_id] = NodeGene(
node_id=new_id,
node_type='hidden',
activation='relu'
)
# Create two new connections
self.connection_genes.extend([
ConnectionGene(
source=source,
target=new_id,
weight=1.0,
enabled=True,
innovation=self.innovation_number
),
ConnectionGene(
source=new_id,
target=target,
weight=conn.weight,
enabled=True,
innovation=self.innovation_number + 1
)
])
self.innovation_number += 2
break
def add_connection(self, source: int, target: int, weight: Optional[float] = None) -> bool:
"""Add a new connection between two nodes.
Args:
source: Source node ID
target: Target node ID
weight: Optional connection weight. If None, a random weight is generated.
Returns:
True if connection was added, False if invalid or already exists
"""
# Check if connection already exists
if any(c.source == source and c.target == target for c in self.connection_genes):
return False
# Validate nodes exist
if source not in self.node_genes or target not in self.node_genes:
return False
# Ensure feed-forward (no cycles)
if source >= target: # Simple way to ensure feed-forward
return False
# Generate random weight if not provided
if weight is None:
weight = float(jrandom.normal(self.key) * 1.0)
# Add new connection
self.connection_genes.append(ConnectionGene(
source=source,
target=target,
weight=weight,
enabled=True,
innovation=self.innovation_number
))
self.innovation_number += 1
return True
def crossover(self, other: 'Genome', key: jnp.ndarray) -> 'Genome':
"""Perform crossover between two genomes.
Args:
other: Other parent genome
key: JAX PRNG key
Returns:
Child genome
"""
# Create child genome
child = Genome(self.input_size, self.output_size)
# Inherit node genes
for node_id in self.node_genes:
if node_id in other.node_genes:
# Inherit randomly from either parent
if jrandom.uniform(key) < 0.5:
child.node_genes[node_id] = self.node_genes[node_id]
else:
child.node_genes[node_id] = other.node_genes[node_id]
else:
# Inherit from fitter parent
child.node_genes[node_id] = self.node_genes[node_id]
# Inherit connection genes
for conn in self.connection_genes:
if conn.innovation in [c.innovation for c in other.connection_genes]:
# Inherit randomly from either parent
other_conn = next(c for c in other.connection_genes if c.innovation == conn.innovation)
if jrandom.uniform(key) < 0.5:
child.connection_genes.append(ConnectionGene(
source=conn.source,
target=conn.target,
weight=conn.weight,
enabled=conn.enabled,
innovation=conn.innovation
))
else:
child.connection_genes.append(ConnectionGene(
source=other_conn.source,
target=other_conn.target,
weight=other_conn.weight,
enabled=other_conn.enabled,
innovation=other_conn.innovation
))
else:
# Inherit from fitter parent
child.connection_genes.append(ConnectionGene(
source=conn.source,
target=conn.target,
weight=conn.weight,
enabled=conn.enabled,
innovation=conn.innovation
))
return child
def clone(self) -> 'Genome':
"""Create a copy of this genome.
Returns:
Copy of genome
"""
clone = Genome(self.input_size, self.output_size)
clone.node_genes = self.node_genes.copy()
clone.connection_genes = [ConnectionGene(**conn.__dict__) for conn in self.connection_genes]
return clone
@property
def n_nodes(self) -> int:
"""Get total number of nodes in the genome."""
return len(self.node_genes)