Final files

Abschlussprojekt/connection.py

@@ -0,0 +1,16 @@
+import dataclasses
+
+
+@dataclasses.dataclass()
+class ConnectionGene:
+    nodes: tuple[int, int]
+    weight: float
+    innovation_no: int
+    disabled: bool = False
+
+
+def reset_innovation_numbers():
+    _CONNECTION_GENES.clear()
+
+
+_CONNECTION_GENES: dict[tuple[int, int], ConnectionGene] = dict()

Abschlussprojekt/genome.py

@@ -1,17 +1,20 @@
 from __future__ import annotations
 
-from dataclasses import dataclass
+import dataclasses
+import itertools
 from enum import Enum
 from random import choice
+
+import matplotlib.pyplot as plt
 import numpy as np
+
 from graphs import creates_cycle
 
 rng = np.random.default_rng()
 
-class NodeType(Enum):
-    INPUT = 1
-    HIDDEN = 2
-    OUTPUT = 3
+
+from connection import _CONNECTION_GENES, ConnectionGene
+from node import NodeGene, NodeType
 
 
 class MutationType(Enum):
@@ -19,23 +22,6 @@ class MutationType(Enum):
     ADD_NODE = 2
 
 
-@dataclass(frozen=True)
-class NodeGene:
-    id: int
-    type: NodeType
-
-
-@dataclass()
-class ConnectionGene:
-    nodes: tuple[int, int]
-    weight: float
-    innovation_no: int
-    disabled: bool = False
-
-
-_CONNECTION_GENES: dict[tuple[int, int], ConnectionGene] = dict()
-
-
 class Genome:
     def __init__(self):
         # Initialize nodes
@@ -44,6 +30,14 @@ class Genome:
         # Initialize connections
         self.connections: dict[tuple[int, int], ConnectionGene] = dict()
 
+        self.fitness = 0
+
+    def set_node(self, key: int, node: NodeGene) -> None:
+        self.nodes[key] = node
+
+    def set_connection(self, key: tuple[int, int], connection: ConnectionGene) -> None:
+        self.connections[key] = connection
+
     def add_node(self, node_type: NodeType = NodeType.HIDDEN) -> int:
         """
         Adds a node of the given type to the genome and returns the identification key.
@@ -58,6 +52,11 @@ class Genome:
         Adds a connection of weight between two given nodes to the genome and returns
         the identification key.
         """
+        if not isinstance(from_node, int) or not isinstance(to_node, int):
+            raise ValueError("Nodes must be integer keys.")
+
+        if from_node not in self.nodes or to_node not in self.nodes:
+            raise ValueError("Nodes do not exist.")
 
         key = (from_node, to_node)
         connection = ConnectionGene(key, weight, -1)
@@ -90,34 +89,103 @@ class Genome:
 
         return genome
 
+    @staticmethod
+    def copy(genome: Genome) -> Genome:
+        clone = Genome()
+
+        # Copy nodes
+        for key, node in genome.nodes.items():
+            clone.set_node(key, dataclasses.replace(node))
+
+        # Copy connections
+        for key, connection in genome.connections.items():
+            clone.set_connection(key, dataclasses.replace(connection))
+
+        # Set fitness
+        clone.fitness = genome.fitness
+
+        return clone
+
 
 def mutate(genome: Genome) -> None:
     mutation = choice([MutationType.ADD_NODE, MutationType.ADD_CONNECTION])
-    print(mutation)
     if mutation is MutationType.ADD_CONNECTION:
         _mutate_add_connection(genome)
     elif mutation is MutationType.ADD_NODE:
         _mutate_add_node(genome)
 
 
+def crossover(mother: Genome, father: Genome) -> Genome:
+    mother_connections = {conn.innovation_no: conn for conn in mother.connections.values()}
+    father_connections = {conn.innovation_no: conn for conn in father.connections.values()}
+    innovation_numbers = set(mother_connections.keys()) | set(father_connections.keys())
+
+    child_connections: dict[int, ConnectionGene] = {}
+
+    for i in innovation_numbers:
+        # Matching genes
+        if i in mother_connections and i in father_connections:
+            child_connections[i] = choice((mother_connections[i], father_connections[i]))
+
+        # Disjoint or excess
+        else:
+            # Mother has better fitness
+            if mother.fitness > father.fitness and i in mother_connections:
+                child_connections[i] = mother_connections[i]
+
+            # Father has better fitness
+            elif father.fitness > mother.fitness and i in father_connections:
+                child_connections[i] = father_connections[i]
+
+            # Equal fitness
+            else:
+                connection = choice((mother_connections.get(i, None), father_connections.get(i, None)))
+                if connection is not None:
+                    child_connections[i] = connection
+
+    # Determine input/output dimensions
+    inputs = sum(node.type == NodeType.INPUT for node in mother.nodes.values())
+    outputs = sum(node.type == NodeType.OUTPUT for node in mother.nodes.values())
+
+    # Create child and set nodes & connections
+    child = Genome.new(inputs, outputs)
+    for connection in child_connections.values():
+        # Set connections
+        child.set_connection(connection.nodes, dataclasses.replace(connection))
+        from_node, to_node = connection.nodes
+
+        # Add nodes if required
+        if from_node not in child.nodes:
+            child.set_node(from_node, NodeGene(from_node, NodeType.HIDDEN))
+        if to_node not in child.nodes:
+            child.set_node(to_node, NodeGene(to_node, NodeType.HIDDEN))
+
+    return child
+
+
 def _mutate_add_connection(genome: Genome) -> None:
     """
     In the add_connection mutation, a single new connection gene with a random weight
     is added connecting two previously unconnected nodes.
     """
 
-    from_node = choice([node for node in genome.nodes.values() if not node.type != NodeType.OUTPUT])
-    to_node = choice([node for node in genome.nodes.values() if node.type != NodeType.INPUT])
-
-    # Checking if connection already exists
-    if (from_node, to_node) in genome.connections:
+    from_node = choice([id for id, node in genome.nodes.items() if node.type != NodeType.OUTPUT])
+    try:
+        to_node = choice(
+            [
+                id
+                for id, node in genome.nodes.items()
+                if node.type != NodeType.INPUT and (from_node, id) not in genome.connections
+            ]
+        )
+    except IndexError:
         return
 
     # Checking for cycles
-    if creates_cycle(genome.connections.keys(), (from_node.id, to_node.id)):
+    if creates_cycle(genome.connections.keys(), (from_node, to_node)):
         return
 
-    genome.add_connection(from_node, to_node, weight=rng.uniform(0,1))
+    genome.add_connection(from_node, to_node, weight=rng.uniform(0, 1))
 
 
 def _mutate_add_node(genome: Genome) -> None:
@@ -145,3 +213,68 @@ def _mutate_add_node(genome: Genome) -> None:
 
     # Connection new_node to previous to_node
     genome.add_connection(new_node, to_node, weight=connection.weight)
+
+
+def _excess(g1: Genome, g2: Genome) -> list[int]:
+    g1_connections = {conn.innovation_no: conn for conn in g1.connections.values()}
+    g2_connections = {conn.innovation_no: conn for conn in g2.connections.values()}
+
+    less_connections, more_connections = sorted((g1_connections, g2_connections), key=lambda c: max(c.keys()))
+    return [k for k in more_connections.keys() if k > max(less_connections.keys())]
+
+
+def _disjoint(g1: Genome, g2: Genome) -> list[int]:
+    g1_connections = {conn.innovation_no: conn for conn in g1.connections.values()}
+    g2_connections = {conn.innovation_no: conn for conn in g2.connections.values()}
+
+    less_connections, more_connections = sorted((g1_connections, g2_connections), key=lambda c: max(c.keys()))
+    return list(
+        {i for i in less_connections.keys() if i not in more_connections}
+        | {i for i in more_connections.keys() if i not in less_connections and i <= max(less_connections.keys())}
+    )
+
+
+def _get_delta(g1: Genome, g2: Genome, c1: float, c2: float, c3: float) -> float:
+    n = max([len(g1.nodes), len(g2.nodes)])
+
+    g1_connections = {conn.innovation_no: conn for conn in g1.connections.values()}
+    g2_connections = {conn.innovation_no: conn for conn in g2.connections.values()}
+    innovation_numbers = set(g1_connections.keys()) | set(g2_connections.keys())
+
+    # Calculate number of excess genes
+    less_connections, more_connections = sorted((g1_connections, g2_connections), key=lambda c: max(c.keys()))
+    e = len([k for k in more_connections.keys() if k > max(less_connections.keys())])
+
+    # Calculate number of disjoint genes
+    d = len(
+        {i for i in less_connections.keys() if i not in more_connections}
+        | {i for i in more_connections.keys() if i not in less_connections and i <= max(less_connections.keys())}
+    )
+
+    # Average weight difference of matching genes
+    w = 0
+    for i in innovation_numbers:
+        if i in g1_connections and i in g2_connections:
+            w += abs(g1_connections[i].weight - g2_connections[i].weight)
+
+    delta = ((c1 * e) / n) + ((c2 * d) / n) + (c3 * w)
+    return delta
+
+
+def specify(genomes: list, c1: float, c2: float, c3: float) -> list[list]:
+    THRESHOLD = 1
+    species = []
+    for genom in genomes:
+        done = False
+        if len(species) < 1:
+            species.append([genom])
+            done = True
+        for spicy in species:
+            print("Delta: ", _get_delta(genom, spicy[0], c1, c2, c3))
+            if _get_delta(genom, spicy[0], c1, c2, c3) < THRESHOLD and not done:
+                spicy.append(genom)
+                done = True
+        if not done:
+            species.append([genom])
+
+    return species

Abschlussprojekt/node.py

@@ -0,0 +1,14 @@
+import dataclasses
+from enum import Enum
+
+
+class NodeType(Enum):
+    INPUT = 1
+    HIDDEN = 2
+    OUTPUT = 3
+
+
+@dataclasses.dataclass(frozen=True)
+class NodeGene:
+    id: int
+    type: NodeType

Abschlussprojekt/visualization.py

@@ -1,10 +1,10 @@
-import itertools
-
 import matplotlib.pyplot as plt
 import networkx as nx
 import numpy as np
+import tabulate
 
 from genome import Genome, NodeType, mutate
+from node import NodeType
 
 
 def _find_layer(g: nx.DiGraph, hidden_node: int, inputs: list[int]) -> int:
@@ -13,10 +13,10 @@ def _find_layer(g: nx.DiGraph, hidden_node: int, inputs: list[int]) -> int:
         paths += list(nx.all_simple_paths(g, input_node, hidden_node))
 
     path_lengths = [len(path) for path in paths]
-    return max(path_lengths)
+    return 2 if len(path_lengths) == 0 else max(path_lengths)
 
 
-def genome(genome: Genome):
+def genome_graph(genome: Genome):
     graph = nx.DiGraph()
 
     # Add nodes
@@ -49,19 +49,35 @@ def genome(genome: Genome):
 
     plt.subplot()
     pos = nx.multipartite_layout(graph, subset_key="layer")
-    nx.draw_networkx_nodes(graph, pos, nodelist=inputs, label=inputs, node_color="#ff0000")
-    nx.draw_networkx_nodes(graph, pos, nodelist=hidden, label=hidden, node_color="#00ff00")
-    nx.draw_networkx_nodes(graph, pos, nodelist=outputs, label=outputs, node_color="#0000ff")
+    nx.draw_networkx_nodes(graph, pos, nodelist=inputs, node_color="#ff0000")
+    nx.draw_networkx_nodes(graph, pos, nodelist=hidden, node_color="#00ff00")
+    nx.draw_networkx_nodes(graph, pos, nodelist=outputs, node_color="#0000ff")
+    nx.draw_networkx_labels(graph, pos)
     nx.draw_networkx_edges(graph, pos)
 
 
+def genome_table(genome: Genome):
+    table = [
+        (conn.innovation_no, "->".join([str(n) for n in conn.nodes]), "DIS" if conn.disabled else "")
+        for conn in genome.connections.values()
+    ]
+    table.sort(key=lambda c: c[0])
+    table = zip(*table)
+
+    print(tabulate.tabulate(table, tablefmt="psql"))
+
+
 if __name__ == "__main__":
     g1 = Genome.new(3, 2)
     g1.add_connection(0, 4, 0.5)
     mutate(g1)
     mutate(g1)
     mutate(g1)
+    mutate(g1)
+    mutate(g1)
+    mutate(g1)
 
     # mutate(g1)
-    genome(g1)
-    plt.show()
+    # genome_graph(g1)
+    # plt.show()
+    genome_table(g1)