Sam
3b64ef62e1
Some checks failed
Build Simulation and Test / Run All Tests (push) Failing after 2m35s
467 lines
20 KiB
Python
467 lines
20 KiB
Python
import numpy as np
|
|
import random
|
|
from copy import deepcopy
|
|
|
|
|
|
class FlexibleNeuralNetwork:
|
|
"""
|
|
A flexible neural network that can mutate its structure and weights.
|
|
Supports variable topology with cross-layer connections.
|
|
"""
|
|
|
|
def __init__(self, input_size=2, output_size=2, empty_start=True):
|
|
self.input_size = input_size
|
|
self.output_size = output_size
|
|
|
|
# Network structure: list of layers, each layer is a list of neurons
|
|
# Each neuron is represented by its connections and bias
|
|
self.layers = []
|
|
|
|
# Initialize network based on empty_start parameter
|
|
if empty_start:
|
|
self._initialize_empty_network()
|
|
else:
|
|
self._initialize_basic_network()
|
|
|
|
self.network_cost = self.calculate_network_cost()
|
|
|
|
def _initialize_basic_network(self):
|
|
"""Initialize a basic network with input->output connections only."""
|
|
# Input layer (no actual neurons, just placeholders)
|
|
input_layer = [{'type': 'input', 'id': i} for i in range(self.input_size)]
|
|
|
|
# Output layer with connections to all inputs
|
|
output_layer = []
|
|
for i in range(self.output_size):
|
|
neuron = {
|
|
'type': 'output',
|
|
'id': f'out_{i}',
|
|
'bias': random.uniform(-1, 1),
|
|
'connections': [] # List of (source_layer, source_neuron, weight)
|
|
}
|
|
|
|
# Connect to all input neurons
|
|
for j in range(self.input_size):
|
|
neuron['connections'].append((0, j, random.uniform(-2, 2)))
|
|
|
|
output_layer.append(neuron)
|
|
|
|
self.layers = [input_layer, output_layer]
|
|
|
|
def _initialize_empty_network(self):
|
|
"""Initialize an empty network with no connections or biases."""
|
|
# Input layer (no actual neurons, just placeholders)
|
|
input_layer = [{'type': 'input', 'id': i} for i in range(self.input_size)]
|
|
|
|
# Output layer with no connections and zero bias
|
|
output_layer = []
|
|
for i in range(self.output_size):
|
|
neuron = {
|
|
'type': 'output',
|
|
'id': f'out_{i}',
|
|
'bias': 0.0,
|
|
'connections': [] # Empty connections list
|
|
}
|
|
output_layer.append(neuron)
|
|
|
|
self.layers = [input_layer, output_layer]
|
|
|
|
def _remove_duplicate_connections(self):
|
|
"""Remove duplicate connections and keep only the last weight for each unique connection."""
|
|
for layer in self.layers[1:]: # Skip input layer
|
|
for neuron in layer:
|
|
if 'connections' not in neuron:
|
|
continue
|
|
|
|
# Use a dictionary to track unique connections by (source_layer, source_neuron)
|
|
unique_connections = {}
|
|
|
|
for source_layer, source_neuron, weight in neuron['connections']:
|
|
connection_key = (source_layer, source_neuron)
|
|
# Keep the last weight encountered for this connection
|
|
unique_connections[connection_key] = weight
|
|
|
|
# Rebuild connections list without duplicates
|
|
neuron['connections'] = [
|
|
(source_layer, source_neuron, weight)
|
|
for (source_layer, source_neuron), weight in unique_connections.items()
|
|
]
|
|
|
|
def _connection_exists(self, target_neuron, source_layer_idx, source_neuron_idx):
|
|
"""Check if a connection already exists between two neurons."""
|
|
if 'connections' not in target_neuron:
|
|
return False
|
|
|
|
for source_layer, source_neuron, weight in target_neuron['connections']:
|
|
if source_layer == source_layer_idx and source_neuron == source_neuron_idx:
|
|
return True
|
|
return False
|
|
|
|
def forward(self, inputs):
|
|
"""
|
|
Forward pass through the network.
|
|
|
|
:param inputs: List or array of input values
|
|
:return: List of output values
|
|
"""
|
|
if len(inputs) != self.input_size:
|
|
raise ValueError(f"Expected {self.input_size} inputs, got {len(inputs)}")
|
|
|
|
# Store activations for each layer
|
|
activations = [inputs] # Input layer activations
|
|
|
|
# Process each subsequent layer
|
|
for layer_idx in range(1, len(self.layers)):
|
|
layer_activations = []
|
|
|
|
for neuron in self.layers[layer_idx]:
|
|
if neuron['type'] == 'input':
|
|
continue # Skip input neurons in hidden layers
|
|
|
|
# Calculate weighted sum of inputs
|
|
weighted_sum = 0.0 # Start with 0 instead of bias
|
|
|
|
# Only add bias if neuron has connections
|
|
if 'connections' in neuron and len(neuron['connections']) > 0:
|
|
weighted_sum = neuron['bias']
|
|
|
|
for source_layer, source_neuron, weight in neuron['connections']:
|
|
if source_layer < len(activations):
|
|
if source_neuron < len(activations[source_layer]):
|
|
weighted_sum += activations[source_layer][source_neuron] * weight
|
|
|
|
# Apply activation function (tanh for bounded output)
|
|
# If no connections and no bias applied, this will be tanh(0) = 0
|
|
activation = np.tanh(weighted_sum)
|
|
layer_activations.append(activation)
|
|
|
|
activations.append(layer_activations)
|
|
|
|
return activations[-1] # Return output layer activations
|
|
|
|
def mutate(self, mutation_rate=0.1):
|
|
"""
|
|
Create a mutated copy of this network.
|
|
|
|
:param mutation_rate: Base probability multiplied by specific mutation weights
|
|
:return: New mutated FlexibleNeuralNetwork instance
|
|
"""
|
|
mutated = deepcopy(self)
|
|
|
|
# Weighted mutations (probability = mutation_rate * weight)
|
|
# Higher weights = more likely to occur
|
|
mutations = [
|
|
(mutated._mutate_weights, 5.0), # Most common - fine-tune existing
|
|
(mutated._mutate_biases, 3.0), # Common - adjust neuron thresholds
|
|
(mutated._add_connection, 1.5), # Moderate - grow connectivity
|
|
(mutated._remove_connection, 0.8), # Less common - reduce connectivity
|
|
(mutated._add_neuron, 0.3), # Rare - structural growth
|
|
(mutated._remove_neuron, 0.1), # Very rare - structural reduction
|
|
(mutated._add_layer, 0.05), # New: create a new layer (very rare)
|
|
]
|
|
|
|
# Apply weighted random mutations
|
|
for mutation_func, weight in mutations:
|
|
if random.random() < (mutation_rate * weight):
|
|
mutation_func()
|
|
|
|
# Clean up any duplicate connections that might have been created
|
|
mutated._remove_duplicate_connections()
|
|
|
|
# Ensure the network maintains basic connectivity
|
|
mutated._ensure_network_connectivity()
|
|
|
|
mutated.network_cost = mutated.calculate_network_cost()
|
|
|
|
return mutated
|
|
|
|
def _mutate_weights(self):
|
|
"""Slightly modify existing connection weights."""
|
|
for layer in self.layers[1:]: # Skip input layer
|
|
for neuron in layer:
|
|
if 'connections' in neuron:
|
|
for i in range(len(neuron['connections'])):
|
|
if random.random() < 0.3: # 30% chance to mutate each weight
|
|
source_layer, source_neuron, weight = neuron['connections'][i]
|
|
# Add small random change
|
|
new_weight = weight + random.uniform(-0.5, 0.5)
|
|
neuron['connections'][i] = (source_layer, source_neuron, new_weight)
|
|
|
|
def _mutate_biases(self):
|
|
"""Slightly modify neuron biases."""
|
|
for layer in self.layers[1:]: # Skip input layer
|
|
for neuron in layer:
|
|
if 'bias' in neuron and random.random() < 0.3:
|
|
neuron['bias'] += random.uniform(-0.5, 0.5)
|
|
|
|
def _add_connection(self):
|
|
"""Add a new random connection."""
|
|
if len(self.layers) < 2:
|
|
return
|
|
|
|
# Find layers with neurons
|
|
valid_target_layers = []
|
|
for i in range(1, len(self.layers)):
|
|
if len(self.layers[i]) > 0:
|
|
valid_target_layers.append(i)
|
|
|
|
if not valid_target_layers:
|
|
return
|
|
|
|
# Pick a random target neuron (not in input layer)
|
|
target_layer_idx = random.choice(valid_target_layers)
|
|
target_neuron_idx = random.randint(0, len(self.layers[target_layer_idx]) - 1)
|
|
target_neuron = self.layers[target_layer_idx][target_neuron_idx]
|
|
|
|
if 'connections' not in target_neuron:
|
|
return
|
|
|
|
# Find valid source layers (must have neurons and be before target)
|
|
valid_source_layers = []
|
|
for i in range(target_layer_idx):
|
|
if len(self.layers[i]) > 0:
|
|
valid_source_layers.append(i)
|
|
|
|
if not valid_source_layers:
|
|
return
|
|
|
|
# Pick a random source (from any previous layer with neurons)
|
|
source_layer_idx = random.choice(valid_source_layers)
|
|
source_neuron_idx = random.randint(0, len(self.layers[source_layer_idx]) - 1)
|
|
|
|
# Check if connection already exists using the helper method
|
|
if self._connection_exists(target_neuron, source_layer_idx, source_neuron_idx):
|
|
return # Connection already exists, don't add duplicate
|
|
|
|
# Add new connection
|
|
new_weight = random.uniform(-2, 2)
|
|
target_neuron['connections'].append((source_layer_idx, source_neuron_idx, new_weight))
|
|
|
|
def _remove_connection(self):
|
|
"""Remove a random connection."""
|
|
for layer in self.layers[1:]:
|
|
for neuron in layer:
|
|
if 'connections' in neuron and len(neuron['connections']) > 1:
|
|
if random.random() < 0.1: # 10% chance to remove a connection
|
|
neuron['connections'].pop(random.randint(0, len(neuron['connections']) - 1))
|
|
|
|
def _add_neuron(self):
|
|
"""Add a new neuron to a random hidden layer or create a new hidden layer."""
|
|
if len(self.layers) == 2: # Only input and output layers
|
|
# Create a new hidden layer
|
|
hidden_neuron = {
|
|
'type': 'hidden',
|
|
'id': f'hidden_{random.randint(1000, 9999)}',
|
|
'bias': random.uniform(-1, 1),
|
|
'connections': []
|
|
}
|
|
|
|
# Connect to some input neurons (avoid duplicates)
|
|
for i in range(self.input_size):
|
|
if random.random() < 0.7: # 70% chance to connect to each input
|
|
if not self._connection_exists(hidden_neuron, 0, i):
|
|
hidden_neuron['connections'].append((0, i, random.uniform(-2, 2)))
|
|
|
|
# Insert hidden layer
|
|
self.layers.insert(1, [hidden_neuron])
|
|
|
|
# Update output layer connections to potentially use new hidden neuron
|
|
for neuron in self.layers[-1]: # Output layer (now at index 2)
|
|
if random.random() < 0.5: # 50% chance to connect to new hidden neuron
|
|
if not self._connection_exists(neuron, 1, 0):
|
|
neuron['connections'].append((1, 0, random.uniform(-2, 2)))
|
|
|
|
else:
|
|
# Add neuron to existing hidden layer
|
|
# Find hidden layers that exist
|
|
hidden_layer_indices = []
|
|
for i in range(1, len(self.layers) - 1):
|
|
if i < len(self.layers): # Safety check
|
|
hidden_layer_indices.append(i)
|
|
|
|
if not hidden_layer_indices:
|
|
return
|
|
|
|
hidden_layer_idx = random.choice(hidden_layer_indices)
|
|
new_neuron = {
|
|
'type': 'hidden',
|
|
'id': f'hidden_{random.randint(1000, 9999)}',
|
|
'bias': random.uniform(-1, 1),
|
|
'connections': []
|
|
}
|
|
|
|
# Connect to some neurons from previous layers (avoid duplicates)
|
|
for layer_idx in range(hidden_layer_idx):
|
|
if len(self.layers[layer_idx]) > 0: # Only if layer has neurons
|
|
for neuron_idx in range(len(self.layers[layer_idx])):
|
|
if random.random() < 0.3: # 30% chance to connect
|
|
if not self._connection_exists(new_neuron, layer_idx, neuron_idx):
|
|
new_neuron['connections'].append((layer_idx, neuron_idx, random.uniform(-2, 2)))
|
|
|
|
self.layers[hidden_layer_idx].append(new_neuron)
|
|
|
|
# Update connections from later layers to potentially connect to this new neuron
|
|
new_neuron_idx = len(self.layers[hidden_layer_idx]) - 1
|
|
for later_layer_idx in range(hidden_layer_idx + 1, len(self.layers)):
|
|
if len(self.layers[later_layer_idx]) > 0: # Only if layer has neurons
|
|
for neuron in self.layers[later_layer_idx]:
|
|
if random.random() < 0.2: # 20% chance to connect to new neuron
|
|
if not self._connection_exists(neuron, hidden_layer_idx, new_neuron_idx):
|
|
neuron['connections'].append((hidden_layer_idx, new_neuron_idx, random.uniform(-2, 2)))
|
|
|
|
def _remove_neuron(self):
|
|
"""Remove a random neuron from hidden layers."""
|
|
if len(self.layers) <= 2: # No hidden layers
|
|
return
|
|
|
|
# Find hidden layers that have neurons
|
|
valid_hidden_layers = []
|
|
for layer_idx in range(1, len(self.layers) - 1): # Only hidden layers
|
|
if len(self.layers[layer_idx]) > 0:
|
|
valid_hidden_layers.append(layer_idx)
|
|
|
|
if not valid_hidden_layers:
|
|
return
|
|
|
|
# Pick a random hidden layer with neurons
|
|
layer_idx = random.choice(valid_hidden_layers)
|
|
neuron_idx = random.randint(0, len(self.layers[layer_idx]) - 1)
|
|
|
|
# Remove the neuron
|
|
self.layers[layer_idx].pop(neuron_idx)
|
|
|
|
# Remove connections to this neuron from later layers
|
|
for later_layer_idx in range(layer_idx + 1, len(self.layers)):
|
|
for neuron in self.layers[later_layer_idx]:
|
|
if 'connections' in neuron:
|
|
neuron['connections'] = [
|
|
(src_layer, src_neuron, weight)
|
|
for src_layer, src_neuron, weight in neuron['connections']
|
|
if not (src_layer == layer_idx and src_neuron == neuron_idx)
|
|
]
|
|
|
|
# Adjust neuron indices for remaining neurons in the same layer
|
|
for later_layer_idx in range(layer_idx + 1, len(self.layers)):
|
|
for neuron in self.layers[later_layer_idx]:
|
|
if 'connections' in neuron:
|
|
adjusted_connections = []
|
|
for src_layer, src_neuron, weight in neuron['connections']:
|
|
if src_layer == layer_idx and src_neuron > neuron_idx:
|
|
# Adjust index down by 1 since we removed a neuron
|
|
adjusted_connections.append((src_layer, src_neuron - 1, weight))
|
|
else:
|
|
adjusted_connections.append((src_layer, src_neuron, weight))
|
|
neuron['connections'] = adjusted_connections
|
|
|
|
# Remove empty hidden layers to keep network clean
|
|
if len(self.layers[layer_idx]) == 0:
|
|
self.layers.pop(layer_idx)
|
|
|
|
# Adjust all layer indices in connections that reference layers after the removed one
|
|
for layer in self.layers:
|
|
for neuron in layer:
|
|
if 'connections' in neuron:
|
|
adjusted_connections = []
|
|
for src_layer, src_neuron, weight in neuron['connections']:
|
|
if src_layer > layer_idx:
|
|
adjusted_connections.append((src_layer - 1, src_neuron, weight))
|
|
else:
|
|
adjusted_connections.append((src_layer, src_neuron, weight))
|
|
neuron['connections'] = adjusted_connections
|
|
|
|
def _add_layer(self):
|
|
"""Add a new hidden layer at a random position with at least one neuron."""
|
|
if len(self.layers) < 2:
|
|
return # Need at least input and output layers
|
|
|
|
# Choose a position between input and output layers
|
|
insert_idx = random.randint(1, len(self.layers) - 1)
|
|
# Create a new hidden neuron
|
|
new_neuron = {
|
|
'type': 'hidden',
|
|
'id': f'hidden_{random.randint(1000, 9999)}',
|
|
'bias': random.uniform(-1, 1),
|
|
'connections': []
|
|
}
|
|
# Connect to all neurons in the previous layer
|
|
for prev_idx in range(len(self.layers[insert_idx - 1])):
|
|
if random.random() < 0.5:
|
|
new_neuron['connections'].append((insert_idx - 1, prev_idx, random.uniform(-2, 2)))
|
|
# Insert the new layer
|
|
self.layers.insert(insert_idx, [new_neuron])
|
|
# Connect neurons in the next layer to the new neuron
|
|
if insert_idx + 1 < len(self.layers):
|
|
for neuron in self.layers[insert_idx + 1]:
|
|
if 'connections' in neuron and random.random() < 0.5:
|
|
neuron['connections'].append((insert_idx, 0, random.uniform(-2, 2)))
|
|
|
|
def _ensure_network_connectivity(self):
|
|
"""Ensure the network maintains basic connectivity from inputs to outputs."""
|
|
# Check if output neurons have any connections
|
|
output_layer = self.layers[-1]
|
|
|
|
for i, output_neuron in enumerate(output_layer):
|
|
if 'connections' not in output_neuron or len(output_neuron['connections']) == 0:
|
|
# Output neuron has no connections - reconnect to input layer
|
|
for j in range(self.input_size):
|
|
if not self._connection_exists(output_neuron, 0, j):
|
|
output_neuron['connections'].append((0, j, random.uniform(-2, 2)))
|
|
break # Add at least one connection
|
|
|
|
# Ensure at least one path exists from input to output
|
|
if len(self.layers) > 2: # Has hidden layers
|
|
# Check if any hidden neurons are connected to inputs
|
|
has_input_connection = False
|
|
for layer_idx in range(1, len(self.layers) - 1): # Hidden layers
|
|
for neuron in self.layers[layer_idx]:
|
|
if 'connections' in neuron:
|
|
for src_layer, src_neuron, weight in neuron['connections']:
|
|
if src_layer == 0: # Connected to input
|
|
has_input_connection = True
|
|
break
|
|
if has_input_connection:
|
|
break
|
|
if has_input_connection:
|
|
break
|
|
|
|
# If no hidden neuron connects to input, create one
|
|
if not has_input_connection and len(self.layers) > 2:
|
|
first_hidden_layer = self.layers[1]
|
|
if len(first_hidden_layer) > 0:
|
|
first_neuron = first_hidden_layer[0]
|
|
if 'connections' in first_neuron:
|
|
# Add connection to first input
|
|
if not self._connection_exists(first_neuron, 0, 0):
|
|
first_neuron['connections'].append((0, 0, random.uniform(-2, 2)))
|
|
|
|
def get_structure_info(self):
|
|
"""Return information about the network structure."""
|
|
info = {
|
|
'total_layers': len(self.layers),
|
|
'layer_sizes': [len(layer) for layer in self.layers],
|
|
'total_connections': 0,
|
|
'total_neurons': sum(len(layer) for layer in self.layers),
|
|
'network_cost': self.network_cost
|
|
}
|
|
|
|
for layer in self.layers[1:]:
|
|
for neuron in layer:
|
|
if 'connections' in neuron:
|
|
info['total_connections'] += len(neuron['connections'])
|
|
|
|
return info
|
|
|
|
def calculate_network_cost(self):
|
|
"""
|
|
Estimate the computational cost of the network.
|
|
Cost is defined as the total number of connections plus the number of neurons
|
|
(i.e., total multiply-accumulate operations and activations per forward pass).
|
|
"""
|
|
total_connections = 0
|
|
total_neurons = 0
|
|
for layer in self.layers[1:]: # Skip input layer (no computation)
|
|
for neuron in layer:
|
|
total_neurons += 1
|
|
if 'connections' in neuron:
|
|
total_connections += len(neuron['connections'])
|
|
return total_connections + total_neurons |