evonet 0.1.0.dev13__tar.gz → 0.1.0.dev15__tar.gz

{evonet-0.1.0.dev13 → evonet-0.1.0.dev15}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: evonet
-Version: 0.1.0.dev13
+Version: 0.1.0.dev15
 Summary: Evolvable neural network core for integration with EvoLib
 Author-email: EvoLib <evolib@dismail.de>
 License: MIT License

{evonet-0.1.0.dev13 → evonet-0.1.0.dev15}/evonet/core.py

@@ -20,6 +20,7 @@ from evonet.connection import Connection
 from evonet.enums import ConnectionType, NeuronRole, RecurrentKind
 from evonet.layer import Layer
 from evonet.neuron import Neuron
+from evonet.utils import connection_init_value
 
 
 class Nnet:
@@ -102,7 +103,7 @@ class Nnet:
         label: str = "",
         role: NeuronRole = NeuronRole.HIDDEN,
         count: int = 1,
-        connection_init: Literal["random", "zero", "none"] = "random",
+        connection_init: Literal["random", "zero", "near_zero", "none"] = "zero",
         recurrent: Optional[set[RecurrentKind]] = None,
     ) -> list[Neuron]:
         """
@@ -116,9 +117,10 @@ class Nnet:
             role: Role of the neuron (INPUT, HIDDEN, OUTPUT).
             count: Number of neurons to add (default: 1).
             connection_init:
-                "random" – connect with random weights (feedforward + recurrent)
-                "zero"   – connect with weight 0.0 (feedforward + recurrent)
-                "none"   – do not create connections (feedforward + recurrent)
+                "random"    - connect with random weights (feedforward + recurrent)
+                "zero"      - connect with weight 0.0 (feedforward + recurrent)
+                "near_zero" - connect with small random weights (-0.05, 0.05)
+                "none"      - do not create connections (feedforward + recurrent)
             recurrent: Optional recurrent connection types.
 
         Returns:
@@ -144,10 +146,7 @@ class Nnet:
             new_neurons.append(neuron)
 
         # Weights based on init mode
-        weight_map = {"random": None, "zero": 0.0, "none": None}
-        if connection_init not in weight_map:
-            raise ValueError(f"Invalid connection_init: {connection_init}")
-        weight = weight_map[connection_init]
+        weight = connection_init_value(connection_init)
 
         skip_connections = connection_init == "none"
 

{evonet-0.1.0.dev13 → evonet-0.1.0.dev15}/evonet/enums.py

@@ -48,3 +48,21 @@ class RecurrentKind(str, Enum):
     DIRECT = "direct"  # src == dst
     LATERAL = "lateral"  # same layer, src != dst
     INDIRECT = "indirect"  # src_layer > dst_layer (back edge)
+
+
+# Neutral activation functions
+# These functions satisfy f(0) ≈ 0 and are safe for zero-weight initialization.
+NEUTRAL_ACTIVATIONS: list[str] = [
+    "tanh",
+    "relu",
+    "relu_max1",
+    "leaky_relu",
+    "elu",
+    "selu",
+    "linear",
+    "linear_max1",
+    "invert",
+    "swish",
+    "mish",
+    "softsign",
+]

{evonet-0.1.0.dev13 → evonet-0.1.0.dev15}/evonet/mutation.py

@@ -19,6 +19,7 @@ from evonet.connection import Connection
 from evonet.core import Nnet
 from evonet.enums import ConnectionType, NeuronRole, RecurrentKind
 from evonet.neuron import Neuron
+from evonet.utils import connection_init_value
 
 ALL_ACTIVATIONS = "all"
 
@@ -181,14 +182,7 @@ def add_random_connection(
     """
 
     # Weight initialization
-    if connection_init == "zero":
-        weight = 0.0
-    elif connection_init == "random":
-        weight = np.random.randn() * 0.5
-    elif connection_init == "near_zero":
-        weight = np.random.uniform(-0.05, 0.05)
-    else:
-        raise ValueError(f"Unknown init mode: {connection_init}")
+    weight = connection_init_value(connection_init)
 
     # normalize to set[RecurrentKind]
     kinds: set[RecurrentKind] = set()
@@ -265,7 +259,11 @@ def remove_random_connection(net: Nnet) -> None:
     conn.target.incoming.remove(conn)
 
 
-def add_random_neuron(net: Nnet, activations: list[str] | None = None) -> None:
+def add_random_neuron(
+    net: Nnet,
+    activations: list[str] | None = None,
+    connection_init: Literal["zero", "random", "near_zero", "none"] = "zero",
+) -> None:
     """
     Insert a new hidden neuron into a random layer.
 
@@ -293,7 +291,7 @@ def add_random_neuron(net: Nnet, activations: list[str] | None = None) -> None:
         layer_idx=net.layers.index(layer),
         activation=random_function_name(activations),
         role=NeuronRole.HIDDEN,
-        connection_init="random",
+        connection_init=connection_init,
     )
 
 

evonet-0.1.0.dev15/evonet/utils.py

@@ -0,0 +1,28 @@
+# SPDX-License-Identifier: MIT
+
+import numpy as np
+
+
+def connection_init_value(mode: str) -> float | None:
+    """
+    Return an initial weight value for a new connection.
+
+    Args:
+        mode (str): Initialization mode. One of:
+            - "zero": weight = 0.0
+            - "near_zero": weight ~ U(-0.05, 0.05)
+            - "random": weight ~ N(0, 0.5)
+            - "none": no initialization (returns None)
+
+    Returns:
+        float | None: Initial connection weight or None.
+    """
+    if mode == "zero":
+        return 0.0
+    if mode == "near_zero":
+        return np.random.uniform(-0.05, 0.05)
+    if mode == "random":
+        return np.random.randn() * 0.5
+    if mode == "none":
+        return None
+    raise ValueError(f"Unknown connection_init mode: {mode}")

{evonet-0.1.0.dev13 → evonet-0.1.0.dev15}/evonet.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: evonet
-Version: 0.1.0.dev13
+Version: 0.1.0.dev15
 Summary: Evolvable neural network core for integration with EvoLib
 Author-email: EvoLib <evolib@dismail.de>
 License: MIT License

{evonet-0.1.0.dev13 → evonet-0.1.0.dev15}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "evonet"
-version = "0.1.0.dev13"
+version = "0.1.0.dev15"
 description = "Evolvable neural network core for integration with EvoLib"
 authors = [
     { name = "EvoLib", email = "evolib@dismail.de" }